ADE20K

ADE20K is a large-scale image segmentation dataset containing over 27,000 images annotated with 150 semantic categories covering both indoor and outdoor scenes, released by MIT CSAIL in 2017. It provides pixel-level semantic, instance, and part-level annotations enabling training and benchmarking of scene parsing and semantic segmentation models. ADE20K serves as the foundational benchmark for the ImageNet Scene Parsing Challenge and has driven significant advances in dense prediction architectures, with state-of-the-art models achieving 62.8 mIoU by 2022 (BEiT-3).

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:hasPart ai:SemanticAnnotationLayer))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:hasPart ai:InstanceAnnotationLayer))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:hasPart ai:PartLevelAnnotationLayer))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:hasPart ai:TrainValTestSplit))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:hasPart ai:CategoryOntology150))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:hasPart ai:ImageAnnotationInterface))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:hasPart ai:EvaluationMetricProtocol))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:hasPart ai:ExtendedOntology847))

Dependency Relationships

SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:requires ai:DataAnnotation))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:requires ai:GroundTruth))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:requires ai:HumanAnnotators))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:requires ai:QualityControlPipeline))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:requires ai:SUNDatabaseSource))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:requires ai:PlacesSceneTaxonomy))

Capability Relationships

SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:SemanticSegmentationEvaluation))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:PanopticSegmentationTraining))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:InstanceSegmentationBenchmarking))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:SceneParsingResearch))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:FoundationModelEvaluation))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:OpenVocabularySegmentation))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:WeaklySupervisedLearning))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:TransferLearning))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:enables ai:ZeroShotSegmentation))

Implementation Relationships

SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:implements ai:mIoUEvaluationProtocol))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:implements ai:PixelLevelAnnotationScheme))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:implements ai:HierarchicalObjectOntology))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:implements ai:SceneParsingChallenge))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:implements ai:PartObjectAnnotationHierarchy))

Reduction Relationships

SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:reducesTo ai:StandardSegmentationBenchmark))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:reducesTo ai:EvaluationCorpus))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:reducesTo ai:AnnotatedImageCollection))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:reducesTo ai:GroundTruthDataset))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:reducesTo ai:PixelLabelCollection))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:reducesTo ai:BenchmarkLeaderboard))
SubClassOf(ai:ADE20KDataset
  ObjectSomeValuesFrom(ai:reducesTo ai:SceneParsingCorpus))

About

• ADE20K is a large-scale Benchmark Dataset for scene parsing and holistic Scene Understanding, produced at MIT CSAIL by Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. It was formally introduced in the paper “Scene Parsing through ADE20K Dataset” at CVPR 2017 and substantially extended in the IJCV 2019 journal version “Semantic Understanding of Scenes through the ADE20K Dataset.” The corpus emerged from the MIT Places Dataset and SUN Database lineage as a dense pixel annotation overlay specifically designed to fill the gap between image-classification benchmarks and full scene understanding — a gap that had persisted as the primary unsolved problem in Computer Vision benchmarking throughout the early 2010s.

• The fundamental insight driving ADE20K’s design was that scene understanding cannot be reduced to object recognition alone. Where ImageNet supplied image-level labels encoding “what category is this image?” and COCO Dataset supplied per-instance segmentation masks encoding “where is each object instance?”, neither benchmark asked “what is every pixel in this scene, including background materials, architectural elements, and object parts?” ADE20K annotated every pixel in every image without exception — covering background “stuff” regions (sky, road, grass, water, ceiling, floor, wall) alongside foreground “thing” instances (person, car, chair, tree, building), and then going further to annotate parts of objects (wheel of car, leg of chair, head of person, branch of tree) — producing a three-level annotation hierarchy that no prior large-scale dataset had systematically attempted.

• The dataset’s canonical evaluation protocol uses mean Intersection over Union (mIoU), computed per class by dividing the intersection of the predicted mask and the Ground Truth mask by their union, then averaging across all 150 categories. This single-number metric has become the lingua franca for comparing Semantic Segmentation methods on ADE20K. The protocol deliberately weights all 150 classes equally regardless of their frequency in the dataset, meaning a model must handle rare categories (fireplace, escalator, ottoman) as accurately as common ones (building, sky, person) to achieve a high mIoU score — a requirement that prevents models from gaming the metric by specialising on dominant classes and ignoring rare ones.

• The ImageNet Scene Parsing Challenge, hosted annually by MIT CSAIL from 2016 onwards, used ADE20K’s 150-category split as its evaluation corpus and attracted hundreds of competing teams from academia and industry in each of its first three years, driving a rapid succession of architectural innovations in Dense Prediction networks. The challenge format — with online submission, a held-out test set with labels withheld from participants, and a public leaderboard updated in real time — mirrored the ImageNet LSVRC structure that had catalysed the Deep Learning revolution in Image Classification between 2012 and 2017, and produced the same galvanising effect on scene parsing research. Each year’s winning solution became the de facto architectural baseline for the field in subsequent years: PSPNet in 2016, UperNet in 2018, and eventually transformer-based architectures from 2020 onwards.

• Structurally, ADE20K divides its 150 standard categories into two annotation types following the Panoptic Segmentation convention: “things” (countable object instances — person, car, chair, building, tree) receive individual instance masks enabling Instance Segmentation evaluation whereby each distinct object gets a unique identity mask; “stuff” (amorphous materials and surfaces — sky, grass, road, water, wall, floor) receive semantic labels shared across all contiguous pixels of that material type without distinguishing individual instances (the sky is “sky” everywhere, not “sky instance 1” and “sky instance 2”). This dual annotation makes ADE20K the natural and most comprehensive testbed for Panoptic Segmentation, which unifies both task types into a single, complete scene labelling that assigns every pixel to either an instance or a stuff region and leaves no pixel unlabelled.

• The extended ADE20K-847 benchmark uses all 847 object categories that appear in the annotated corpus with sufficient frequency (at least 25 annotated images), providing a substantially harder evaluation challenge that exposes models to rare and visually ambiguous categories absent from the 150-category standard vocabulary: categories such as “skirting board,” “dado rail,” “cornice,” specific types of indoor plants, historical furniture styles, and specialised industrial equipment that occur in the dataset’s indoor scene images but were collapsed into broader categories in the 150-class standard vocabulary. Research in 2024–2026 increasingly uses ADE20K-847 as the primary evaluation target for Open Vocabulary Segmentation because the 150-category standard split is approaching saturation for the strongest supervised Foundation Models, with BEiT-3’s 62.8 mIoU leaving less than 15 percentage points of headroom before the annotation-noise ceiling.

• By 2026, ADE20K remains one of the most widely cited benchmarks in Computer Vision, appearing in the evaluation sections of virtually every major Semantic Segmentation paper and as a standard fine-tuning evaluation in Foundation Models release reports. Despite the emergence of newer large-scale datasets — Mapillary Vistas (66 categories, 25,000 images, street scenes only), OpenImages segmentation (600 categories, 9 million images but noisier annotations), and the COCO panoptic extension — ADE20K occupies a unique ecological niche: it is the only benchmark that simultaneously offers breadth of scene types (spanning both indoor and outdoor, urban and rural, daytime and night-time environments), depth of annotation (semantic, instance, and part layers at full resolution), and a decade-long record of community evaluation enabling direct year-over-year comparison of model quality across the field’s entire Deep Learning era.

  Annotation Methodology and Data Quality

• The Data Annotation pipeline for ADE20K was designed with the explicit goal of producing consistent, hierarchically complete pixel-level labels across a semantically diverse corpus, employing a multi-pass annotation workflow that is substantially more rigorous than crowd-sourced annotation pipelines used in other large-scale datasets. The annotation interface — developed in-house at MIT CSAIL specifically for this project — presented each annotator with the full-resolution image alongside a hierarchical category browser organised by scene type, enabling structured labelling from coarse semantic regions down to fine object parts in successive annotation passes, each building on the previous layer.

• The annotation workflow proceeded in three distinct stages. In the first stage, “stuff” region labelling, annotators painted large semantic regions using polygon and flood-fill tools: architectural surfaces (wall, floor, ceiling), outdoor ground covers (grass, pavement, road, dirt), water bodies, and sky were labelled at full resolution. In the second stage, “thing” instance annotation, annotators drew precise polygonal outlines around every visible countable object and assigned a category label from the full 3,688-category fine-grained ontology — not the 150-category standard vocabulary — ensuring that the annotation preserved maximum detail even when the benchmark would later aggregate categories. In the third stage, object-part annotation, annotators drew sub-polygons for each defined part of annotated objects: for vehicles (wheel, window, door, headlight, bumper, windshield); for persons (head, arm, hand, leg, foot, torso); for furniture (chair seat, chair back, chair leg, drawer, handle). This three-layer structure is what distinguishes ADE20K from all contemporary large-scale segmentation benchmarks.

• Quality control operated at multiple stages throughout the annotation process. Automated consistency checks flagged annotations with implausible spatial distributions (a “sky” region appearing below the horizon line, a “floor” region in the image’s top third), category frequency outliers suggesting category confusion, and segmentation masks with jagged boundaries inconsistent with object contours. Human expert reviewers then examined all flagged annotations, correcting clear errors and triggering second-pass re-annotation for images where ambiguity could not be resolved by the original annotator. Inter-annotator agreement studies on a 5% sample found annotation consistency rates of approximately 85% at the semantic (150-category) level and 70% at the part level — substantially higher than the 60-75% typically reported for crowd-sourced segmentation datasets where annotators are recruited without specialist training and quality-controlled only by majority voting.

• The category ontology was constructed iteratively rather than defined a priori. An initial vocabulary of 500 common object categories was provided to annotators, who were instructed to expand it when they encountered objects not yet in the taxonomy. Over the course of the annotation project, the ontology grew to 3,688 entries, with each new category requiring at least five confirmed example annotations before inclusion in the official taxonomy. The final 150-category standard benchmark vocabulary was constructed by collapsing synonymous and perceptually similar fine-grained categories and selecting the 150 that appeared in at least 25% of training images — a frequency threshold ensuring that models could learn meaningful representations from the training data. The remaining 3,538 categories are preserved in the full annotation files and contribute to the ADE20K-847 extended benchmark, which uses the 847 categories appearing in at least 25 annotated images.

• The resulting annotation volume is extraordinary for a manually curated academic dataset. Roughly 27,000 images, each with complete pixel-level annotation at two or three hierarchical levels, amounts to approximately 2.5 billion annotated pixels across the corpus. Each image averages 29 annotated object instances and multiple part sub-regions per instance-level object, making ADE20K among the most densely annotated scene understanding datasets in existence. This density — far exceeding what crowd-sourcing platforms typically achieve at scale, where annotators often focus only on salient objects and leave background regions unlabelled — is the primary reason ADE20K has maintained its status as a premium benchmark even as datasets with far larger image counts (OpenImages: 9 million images) and more categories (LVIS: 1,200+ categories) have emerged. Quality and density of annotation matter more than raw scale for scene understanding research, and ADE20K’s Human-in-the-Loop quality control pipeline produced annotations that remain the gold standard a decade after the dataset’s creation.

  Architectural Progression on ADE20K

• The history of Semantic Segmentation research from 2016 to 2026 is largely written in ADE20K leaderboard entries, with each successive state-of-the-art result representing a distinct architectural paradigm shift. The progression from 41.68 mIoU at launch to 62.8 mIoU in 2022 — a 21-point improvement over six years — reflects the combined effect of Deep Learning architecture improvements, scale of pre-training data, and the fundamental shift from Convolutional Neural Network backbones to Transformer Architecture and then Foundation Models.

• The PSPNet baseline (2016, 41.68 mIoU) established the foundational design principle that global context aggregation is essential for scene parsing. PSPNet’s ResNet-101 backbone was augmented with a Pyramid Pooling Module that captured contextual features at four spatial granularities (1×1, 2×2, 3×3, and 6×6 pooling bins), producing a rich multi-scale representation that could disambiguate visually identical local patches by their global scene context — distinguishing “road” from “pavement” by understanding the broader urban scene, for example. This design insight proved so influential that some form of multi-scale context pooling appears in virtually every subsequent ADE20K-evaluated architecture.

• The DeepLab v3+ (ECCV 2018) architecture replaced pyramid pooling with Atrous Spatial Pyramid Pooling (ASPP), combining dilated convolutional features at multiple rates (6, 12, 18 dilations) to capture multi-scale context without reducing spatial resolution through pooling. Its encoder-decoder structure additionally used low-level feature skip connections to recover spatial detail lost in the encoder’s downsampling stages, improving boundary accuracy. DeepLab v3+ achieved widespread adoption in production systems and remains a standard baseline in applied Computer Vision projects.

• The self-attention era (OCNet and DANet, 2018–2019, ~45 mIoU) introduced Attention Mechanism-based context pooling that aggregated features across all pixels of the same predicted category (OCNet: object context) or across all spatial positions simultaneously (DANet: dual attention across channel and spatial dimensions). These approaches presaged the full Transformer Architecture transition by demonstrating that global self-attention was superior to fixed-scale pyramid pooling for scene parsing — the key insight that would eventually enable transformer-based models to surpass Convolutional Neural Network approaches on ADE20K.

• The HRNet (IEEE TPAMI 2020, 45.1 mIoU) took a fundamentally different path from attention-based context aggregation, maintaining full spatial resolution throughout the network via parallel processing streams at multiple resolutions, with frequent cross-scale feature fusion. This produced spatially precise segmentation predictions without requiring upsampling post-processing, making HRNet particularly valuable for applications requiring accurate boundary localisation such as robotic grasping and medical Image Segmentation.

• The Transformer Architecture era began decisively with SETR (CVPR 2021, 48.6 mIoU), which replaced the CNN backbone with a plain Vision Transformer (ViT) encoder trained with supervised patch-level classification. SETR demonstrated for the first time that purely attention-based models could exceed Convolutional Neural Network baselines on ADE20K, proving that the global receptive field of self-attention provided advantages over the local receptive field of convolutional operations for scene understanding tasks. The Swin Transformer (ICCV 2021, 53.5 mIoU with Swin-L + UperNet) introduced hierarchical feature maps via shifted-window partitioned self-attention, combining the global context advantage of transformers with the multi-scale feature hierarchy that Convolutional Neural Network architectures had provided — producing a step change of 5+ mIoU points over prior state-of-the-art.

• SegFormer (NeurIPS 2021, 51.8 mIoU with SegFormer-B5) demonstrated that architectural complexity in the decoder was unnecessary for high performance: a simple linear MLP decoder combining features from four transformer encoder stages at different resolutions could match or exceed computationally heavier decoding approaches. SegFormer’s efficiency profile — ranging from SegFormer-B0 (3.8M parameters, 37.4 mIoU) to SegFormer-B5 (84M parameters, 51.8 mIoU) — made it the most widely adopted production segmentation backbone in 2022–2024, deployed in embedded systems, edge computing platforms, and mobile applications where the CNN/transformer trade-off previously made dense prediction impractical.

• From 2022 onwards, the benchmark has been dominated by large-scale pre-trained Foundation Models that treat ADE20K fine-tuning as a downstream Transfer Learning evaluation rather than a primary training objective. Mask2Former (CVPR 2022, 57.7 mIoU) achieved the strongest purely architectural result by applying masked Attention Mechanism — restricting cross-attention to within predicted mask regions rather than globally — enabling efficient universal Image Segmentation that handles semantic, instance, and panoptic tasks with a single model. InternImage (CVPR 2023, 58.9 mIoU) demonstrated that deformable convolutional backbones, where convolution kernel positions adapt to input features rather than being fixed, could match transformer performance at lower computational cost. EVA (CVPR 2023, 61.5 mIoU) and BEiT-3 (CVPR 2023, 62.8 mIoU) achieved the two highest results by combining billion-parameter-scale masked image modelling pre-training with standard UperNet decoders, demonstrating that pre-training data scale and quality — rather than decoder architecture — is the primary determinant of ADE20K performance above 60 mIoU. BEiT-3’s result of 62.8 mIoU is the current state-of-the-art on ADE20K-150 as of mid-2026, a position it has held for three years without being surpassed.

• For the ADE20K-847 open-vocabulary variant, the performance landscape is very different: CLIPSeg achieves approximately 5 mIoU under zero-shot conditions, OVSeg approximately 9.0 mIoU, and FC-CLIP approximately 12–15 mIoU. These numbers — representing the current frontier of Open Vocabulary Segmentation research — indicate that open-vocabulary models perform at roughly 20–25% of the accuracy of fully supervised models when faced with the full 847-category vocabulary, highlighting the substantial remaining challenge in building segmentation systems that can generalise beyond their training category vocabulary to arbitrary semantic concepts described in natural language.

  Components / Architecture

• Image corpus: 20,210 training images, 2,000 validation images, 3,352 test images (total 25,562 in standard split; 27,000+ with later community additions)

• Category taxonomy:

  • 150 standard categories: benchmark evaluation vocabulary
  • 3,688 fine-grained object categories: full annotation ontology
  • 847-category extended benchmark (ADE20K-847): open-vocabulary evaluation

• Annotation layers:

  • Semantic segmentation layer: per-pixel class label from 150-category vocabulary covering every pixel without ignore regions
  • Instance segmentation layer: distinct polygon masks for each countable object instance; enables instance-level evaluation
  • Part segmentation layer: per-pixel labels for object components (wheel, door, leg, window, seat, head, arm, trunk, branch, etc.)

• Annotation tooling: MIT CSAIL hierarchical image parsing interface; enables successive refinement from coarse scene labels to fine object parts in multi-pass workflow

• Source databases: SUN Database scene images (Xiao et al., 2010); MIT Places Dataset scene categories (Zhou et al., 2014)

• Evaluation metric: mean Intersection over Union (mIoU) averaged across all 150 classes; also reports pixel accuracy and per-class IoU profiles

• Standard splits: train (20,210) / val (2,000) with public ground truth; test (3,352) with held-out labels for challenge submission

• Data formats: MATLAB .mat files (original release); JSON (community conversion); COCO-compatible JSON format via community converters; native HuggingFace Datasets format

• Distribution: MIT CSAIL GitHub repository (CSAILVision/ADE20K) with full annotation files; Hugging Face Datasets hub mirror (1aurent/ADE20K); Dataset Ninja mirror

  Use Cases / Major Families

• ADE20K drives research and deployment across six broad application domains, each exploiting different aspects of its annotation richness.

• Autonomous Driving and Robotics:

  • Autonomous Vehicle Navigation stacks use ADE20K-pretrained Semantic Segmentation models as initialisation for proprietary road-scene fine-tuning
  • ADE20K’s broader category vocabulary (vs. Cityscapes’ 19 categories in German street scenes) captures long-tail objects encountered outside city centres: agricultural machinery, construction equipment, unusual road furniture
  • Robotics Perception pipelines for manipulation use ADE20K’s part-level annotations to identify “chair leg” or “cabinet door handle” rather than just the parent object category, enabling grasp point determination
  • ADE20K-benchmarked models are deployed in warehouse robots at Amazon and Ocado for indoor scene navigation and inventory management

• Indoor Scene Understanding:

  • ADE20K’s rich indoor annotation (bedrooms, kitchens, offices, living rooms, industrial spaces) supports models for home robots and accessibility aids
  • Augmented Reality applications for Apple Vision Pro and Meta Quest use semantic segmentation to understand room geometry — identifying walls, floors, ceilings, furniture, and windows — to anchor virtual content and handle occlusion
  • Accessibility technology for visually impaired users relies on ADE20K-benchmarked segmentation for real-time audio environment description, parsing navigable paths, obstacles, and points of interest

• Foundation Model Pre-training and Evaluation: The most important application of ADE20K in the 2022–2026 research landscape is as the canonical transfer-learning evaluation for Foundation Models pre-trained on internet-scale data. Models including CLIP, DINOv2, BEiT-3, InternImage, EVA, and MAE all include ADE20K fine-tuning as a required evaluation in their model release reports, reporting mIoU on the ADE20K-150 validation split alongside ImageNet top-1 accuracy and COCO detection mAP as the standard triplet of visual representation quality benchmarks. The ADE20K mIoU score serves as the primary proxy for spatial semantic representation quality — measuring whether the model has learned to represent not just “what type of scene is this” (ImageNet) but “where is each semantic class at pixel level” (Dense Prediction). A model achieving high ImageNet accuracy but low ADE20K mIoU reveals representations that discriminate object identity without learning pixel-level spatial semantics, suggesting the learned features are spatially coarse and category-discriminative rather than spatially precise and scene-compositional. Conversely, high ADE20K performance from a model with only modest ImageNet accuracy suggests spatial relational pre-training (as in DINOv2’s self-supervised patch-level contrastive learning) is particularly effective at building dense prediction capabilities. This discriminative power makes ADE20K mIoU the single most informative benchmark for evaluating the quality of visual foundation model representations for downstream dense perception tasks in robotics, AR, and autonomous systems.

• Open-Vocabulary and Zero-Shot Segmentation:

  • ADE20K-150 and ADE20K-847 are the primary benchmarks for Open Vocabulary Segmentation and Zero-Shot Learning segmentation research
  • The 847-category split is particularly challenging: visually similar but semantically distinct categories (furniture subtypes, architectural elements, plant species) require fine-grained visual discrimination
  • Zero-shot methods typically achieve 7–20 mIoU on ADE20K-847 compared to 50–62 for fully supervised methods on ADE20K-150, indicating the substantial remaining open-vocabulary gap

• Medical and Scientific Imaging: ADE20K’s hierarchical annotation methodology — particularly its three-level thing/stuff/part hierarchy — directly influenced annotation practices in Medical Imaging segmentation projects, where part-of-object labelling (organ, sub-organ component, lesion within organ tissue type, lesion subregion) mirrors the thing-part annotation hierarchy in structure and purpose. The ADE20K protocol’s insistence on complete coverage of every pixel without ignore regions has also influenced medical segmentation annotation standards, where incomplete labelling (annotating only foreground pathology without labelling background tissue types) had previously made multi-task training difficult. ADE20K-pretrained backbones serve as initialisation for both radiology (CT volumetric, MRI structural, X-ray) and histopathology (whole-slide images at multiple magnification levels) segmentation models, exploiting the dense spatial representations learned from natural scene parsing to provide better weight initialisation than ImageNet classification pretraining for tasks requiring accurate pixel-level boundary localisation — the primary requirement in medical image delineation tasks.

• Scene Graph Generation: ADE20K’s combined instance-level annotations (which establish discrete object identities) and 150-category spatial layouts (which establish the full scene semantic context) together provide an unusually rich semantic grounding resource for Scene Graph generation research — the task of automatically producing structured graph representations of visual scenes where nodes are object instances and edges are spatial or semantic relations. Models that produce subject-predicate-object relation triples (e.g., “person sitting-on chair,” “lamp above table,” “window adjacent-to wall”) need reliable segmentation backends to identify the nodes of the scene graph, and ADE20K-trained segmentation models serve as the scene graph node generator, with the segmentation categories directly mapping to potential graph node types. This connection makes ADE20K downstream-relevant to visual question answering (requiring spatial scene graph traversal to answer “which chair is closer to the door?”), Image Captioning (requiring scene graph serialisation to produce accurate spatial descriptions), and Multimodal AI reasoning tasks requiring grounded spatial understanding — a use case whose importance has grown substantially as multimodal language models are increasingly evaluated on spatial reasoning tasks requiring explicit pixel-level scene understanding rather than holistic image-level pattern matching.

  Academic Context

• ADE20K was introduced in two key publications from MIT CSAIL and has catalysed a decade of architectural innovation in Semantic Segmentation, Panoptic Segmentation, and Scene Understanding research. The dataset’s intellectual lineage runs directly through two prior large-scale visual benchmarks that established the practice of semantically rich scene annotation at scale.

• The CVPR 2017 paper “Scene Parsing through ADE20K Dataset” by Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba presented both the dataset and an associated baseline architecture. The paper articulated the key motivation — that scene understanding requires annotating not just objects but the full spatial context including background regions and object parts — and demonstrated this through quantitative experiments showing that models trained on ADE20K’s richer annotations generalised better to novel scene understanding tasks than models trained on object-centric benchmarks alone. The PSPNet baseline achieved 41.68 mIoU, winning the inaugural ImageNet Scene Parsing Challenge in 2016.

• The IJCV 2019 journal extension “Semantic Understanding of Scenes through the ADE20K Dataset” provided a comprehensive treatment of the annotation methodology, the hierarchical object ontology, challenge results from 2016 to 2018, and extended experiments validating the dataset’s utility for part-level recognition tasks and scene graph generation. Critically, this extended paper made the complete 3,688-category fine-grained annotation files publicly available rather than only the 150-category standard benchmark annotations, enabling a new line of research on fine-grained object recognition, long-tail category learning, and hierarchical part-aware segmentation that had not been possible with prior large-scale benchmarks.

• The intellectual roots of ADE20K lie in two complementary research traditions. The MIT Places Dataset project (Zhou et al., NIPS 2014; extended 2018), created by Bolei Zhou and Antonio Torralba with contributions from Aude Oliva and other MIT colleagues, demonstrated that Convolutional Neural Network models trained on 10 million scene-categorised images learned rich scene-level representations that transferred well to a wide range of downstream scene understanding tasks. The SUN Database (Xiao et al., CVPR 2010), covering 899 scene categories in 130,519 images, established the practice of semantically rich scene labelling and contributed a substantial fraction of the images that were subsequently re-annotated at pixel level for ADE20K. These two projects established MIT CSAIL as the primary academic centre for large-scale scene understanding dataset construction and set the methodological framework that ADE20K extended with dense pixel-level annotation.

• Architectural milestones on ADE20K:

  • PSPNet (Zhao et al., CVPR 2017): Pyramid Pooling Module; 41.68 mIoU — dataset baseline
  • DeepLab v3+ (Chen et al., ECCV 2018): ASPP encoder-decoder; production standard
  • OCNet / DANet (Yuan & Wang 2018, Fu et al. 2019): self-attention context; ~45 mIoU
  • HRNet (Wang et al., IEEE TPAMI 2020): parallel high-resolution streams; 45.1 mIoU
  • SETR (Zheng et al., CVPR 2021): plain ViT encoder; 48.6 mIoU
  • Swin Transformer (Liu et al., ICCV 2021): shifted-window hierarchical ViT; 53.5 mIoU
  • SegFormer (Xie et al., NeurIPS 2021): mix-transformer + MLP decoder; 51.8 mIoU (B5)
  • Mask2Former (Cheng et al., CVPR 2022): masked attention universal segmentation; 57.7 mIoU
  • InternImage (Wang et al., CVPR 2023): deformable convolutions; 58.9 mIoU
  • EVA (Yao et al., CVPR 2023): 1B parameter masked image modelling; 61.5 mIoU
  • BEiT-3 (Wang et al., CVPR 2023): multimodal image-text pre-training; 62.8 mIoU (current SOTA)

• ADE20K-847 open-vocabulary milestones:

  • CLIPSeg (Lüddecke and Ecker, CVPR 2022): CLIP-based promptable segmentation; first major zero-shot ADE20K-847 result
  • OVSeg (Liang et al., CVPR 2023): mask-adapted CLIP; ~9.0 mIoU ADE20K-847
  • FC-CLIP (Yu et al., NeurIPS 2023): frozen CLIP + in-context classification; improved open-vocabulary performance
  • PCA-Seg (2025): revisiting cost aggregation for open-vocabulary part segmentation

• Broader MIT impact: Bolei Zhou’s Network Dissection (Bau et al., CVPR 2017) used ADE20K’s 150-category vocabulary as the concept taxonomy for dissecting Neural Networks unit selectivity; GAN Dissection (Bau et al., ICLR 2019) used ADE20K categories to understand generative model internals; this cross-cutting use established ADE20K’s taxonomy as a foundational semantic vocabulary for computer vision interpretability research.

  Current Landscape (2026)

• As of June 2026, ADE20K retains its position as the primary benchmark for holistic Semantic Segmentation despite operating in a significantly changed research environment characterised by the dominance of Foundation Models, the shift toward Open Vocabulary Segmentation evaluation, and the gradual saturation of the 150-category standard split by the strongest supervised methods. The dataset’s status has evolved from an active competition benchmark (2016–2022) to a canonical transfer-learning evaluation (2022–2026), with its primary function now being the standardised measurement of how well a foundation model’s learned representations transfer to dense pixel-level semantic prediction tasks.

• The ADE20K-150 standard benchmark is approaching saturation from above: the gap between the current state-of-the-art result (62.8 mIoU, BEiT-3) and the theoretical upper bound set by annotation noise and human disagreement (estimated at approximately 75–80 mIoU based on inter-annotator agreement studies) has narrowed to 12–17 mIoU points. This narrowing reduces the benchmark’s discriminative power for comparing the strongest Foundation Models against each other, since small differences in mIoU at this level may reflect training recipe variations rather than fundamental capability differences. Research attention is consequently bifurcating toward ADE20K-847 (the harder 847-category extended benchmark for Open Vocabulary Segmentation), zero-shot and few-shot ADE20K evaluations that test generalisation without full supervision over the standard category vocabulary, and domain-transferred performance assessments measuring how well ADE20K-pretrained models transfer to held-out specialised domains such as satellite imagery, endoscopy video, and industrial inspection.

• Weakly Supervised Learning and Semi-Supervised Learning research using ADE20K as the evaluation benchmark has intensified considerably since 2023, driven by the practical need to reduce Data Annotation costs for domain-specific applications where full pixel annotation is prohibitively expensive. Foundation model-assisted annotation approaches — specifically using SAM-generated pseudo-labels to initialise supervised segmentation model training — have been demonstrated to achieve SegFormer-B4-level performance (approximately 50.3 mIoU on ADE20K-150) without any full pixel annotations, demonstrating that strong pre-trained spatial priors from SAM can substitute for dense human-labelled ground truth in data-rich scenarios where unlabelled images are plentiful. These approaches have reduced annotation cost estimates for ADE20K-scale labelling projects by 60–80%, making it feasible to produce domain-specific benchmark datasets at ADE20K quality levels in resource-constrained settings.

• Multimodal AI systems — large vision-language models trained on image-text pairs — are increasingly evaluated using ADE20K images as probe stimuli for spatial grounding and semantic understanding tasks. Models such as GPT-4V, Claude 3 Opus, Gemini 1.5 Pro, and LLaVA are tested on tasks requiring them to identify and describe specific ADE20K-annotated regions when prompted in natural language (“What material is the floor in this image?” “Which objects in this scene could a person sit on?” “How many windows are visible in this room?”). These evaluations bridge ADE20K’s pixel-level annotation vocabulary to the emergent spatial semantic representations of large Multimodal AI models, probing whether image-text training is sufficient to develop accurate pixel-level spatial understanding without explicit segmentation supervision — a research question with substantial implications for whether future Foundation Models will be able to perform dense prediction tasks without task-specific fine-tuning.

• SAM 2 (Ravi et al., 2024) is routinely evaluated on ADE20K subsets to assess Zero-Shot Learning segmentation generalisation. SAM 2 achieves notably higher zero-shot mIoU than SAM 1 on the 150-category validation split, primarily due to improvements in the mask decoder’s ability to produce coherent, complete segmentation masks from sparse point and bounding box prompts. The evaluation protocol for foundation model zero-shot segmentation on ADE20K typically provides SAM with the ground-truth bounding box of each object as a prompt (simulating an oracle that knows object locations) and measures whether SAM can produce segmentation masks that match the ground-truth pixel annotations — a test of mask quality given location knowledge, rather than detection capability.

• The industrial deployment of ADE20K-pretrained models has broadened significantly through 2024–2026. In Autonomous Vehicle Navigation, Waymo’s Generation 5 sensor suite, Cruise’s autonomous vehicle platform, and Mobileye’s EyeQ computational perception systems all use ADE20K-benchmarked Semantic Segmentation architectures as off-the-shelf pretrained backbones for proprietary road-scene fine-tuning, with the ADE20K pretraining providing stable feature representations for the long tail of unusual objects (construction equipment, emergency vehicles, cyclists with cargo) that are underrepresented in automotive-specific datasets. In Augmented Reality, Apple Vision Pro’s spatial computing platform, Meta Quest’s mixed reality environment, and Microsoft HoloLens 2’s workspace digitalisation all use ADE20K-pretrained SegFormer and Mask2Former weights for real-time scene mesh understanding, room boundary detection, and dynamic occlusion handling of physical objects. In retail and logistics, Amazon Robotics and Ocado Technology deploy ADE20K-pretrained segmentation models for warehouse environment understanding, autonomous shelf scanning, and robot navigation through cluttered pick-and-pack environments where the dense semantic context about floor space, shelving, and product zones is essential for safe robot operation.

  UK Context

• UK academic contributions to ADE20K-benchmarked research are concentrated at leading institutions across the country, with both southern universities and Northern England’s industrial research ecosystem making distinctive contributions to the field.

• The University of Oxford contributes through two complementary research groups. The Visual Geometry Group (VGG), led by Andrew Zisserman and Philip Torr, has published extensively on dense prediction tasks evaluated on ADE20K, with particular focus on cross-domain generalisation — training on ADE20K and evaluating transfer to proprietary domain datasets — and on architecturally efficient models deployable at the inference latencies required by robotics and AR applications. The Computer Vision and Machine Learning group (CVML), co-led by Torr and Víctor Prisacariu, researches transformer-based Semantic Segmentation and semantic consistency regularisation techniques that improve ADE20K fine-tuning sample efficiency, reducing the number of labelled images required to reach a given performance threshold. Oxford’s EPSRC-funded VisualAI programme grant (2020–2025) explicitly targeted dense visual understanding benchmarked against ADE20K and COCO as its primary evaluation standards, funding foundational work on scene representation, depth-aware segmentation, and video semantic consistency.

• Imperial College London’s Visual Computing group, based in the Department of Computing, has published hardware-aware Semantic Segmentation models targeting embedded deployment on robotics and embedded AI platforms. The 2024 SMOF paper (“Streaming Modern CNNs on FPGAs with Smart Off-Chip”) and HASS paper (“Hardware-aware Sparsity Search”) both used ADE20K mIoU preservation as the primary quality metric for compressed and hardware-optimised dense prediction models, demonstrating that aggressively sparse transformer models can maintain competitive ADE20K performance while reducing inference latency by 3–5× on FPGA hardware — a result directly relevant to industrial robotics where compute budgets are constrained.

• University College London’s Machine Vision group has published research on unified Panoptic Segmentation architectures evaluated simultaneously on ADE20K and COCO Dataset, exploring how a single model can handle both thing-instance and stuff-region labelling in a single forward pass without dedicated heads for each task — a research direction directly motivated by ADE20K’s dual thing-stuff annotation structure. UCL’s interdisciplinary AI Centre also hosts research on fairness and robustness in dense prediction, using ADE20K’s category diversity to probe whether segmentation models perform equitably across scene types and demographic contexts.

• The University of Edinburgh’s VICO group (Visual Interpretation and Computational Ontologies) published “Multi-task Learning with 3D-Aware Regularisation” at ICLR 2024, integrating depth estimation and Semantic Segmentation in a unified 3D-aware training framework with ADE20K as the primary segmentation evaluation benchmark. This work connects ADE20K’s 2D pixel labelling to 3D scene structure, a trajectory directly relevant to Edinburgh’s broader programme on robotic Scene Understanding for navigation in cluttered indoor environments — precisely the indoor scene diversity for which ADE20K’s annotation is most valuable.

• Northern England institutions contribute through applied and industrial AI research programmes. The National Robotics Innovation Centre at the University of Sheffield, co-funded by Innovate UK and UKRI, uses ADE20K-trained segmentation models as the scene perception component in manufacturing robotics research, where industrial scenes require parsing machinery components, workpieces at various assembly stages, safety zones, and human co-workers. Sheffield’s Henry Royce Institute additionally uses ADE20K-style hierarchical annotation methodology for scanning electron microscopy and metallographic image segmentation, applying the dataset’s thing-part hierarchy to microscale scientific imaging. The University of Leeds School of Computing applies ADE20K-benchmarked architectures to agricultural robotics under the Agri-EPI Centre and BBSRC-funded Smart Crop Protection programme, parsing crop rows, soil surface conditions, weed density, and mechanical components in outdoor field environments — a use case where ADE20K’s outdoor scene diversity provides useful pretrained representations even for environments very different from the training distribution. The University of Manchester’s Department of Computer Science participates in the Greater Manchester Combined Authority’s AI for Retail and Logistics initiative, deploying ADE20K-pretrained segmentation models for warehouse inventory management systems where spatial Scene Understanding of shelving, floor areas, and robotic operating zones is required to enable safe human-robot collaboration. The Alan Turing Institute in London has hosted fellowships focused on scene understanding evaluation methodology, producing work on ADE20K annotation quality assessment and initial designs for video and 3D modality extensions that the community will need as 2D static image segmentation approaches saturation.

  Future Directions (2026–2030)

• The primary evolution expected for ADE20K as a research resource runs along five trajectories that will collectively expand its scope from static 2D image benchmarking to dynamic, multimodal, and embodied scene understanding evaluation, and will ultimately reposition the dataset from an active competition benchmark to a foundational calibration tool for increasingly capable Foundation Models.

• Video extension: Temporal consistency in Semantic Segmentation — assigning coherent, time-stable semantic labels to pixels representing the same physical object across consecutive video frames, even as viewpoint, lighting, and object pose change — is the natural next frontier after static image segmentation. An ADE20K-Video extension analogous to the Cityscapes video benchmark but spanning ADE20K’s broader indoor and outdoor scene variety would drive temporal Scene Understanding research, providing training and evaluation data for video-aware segmentation models that must maintain label consistency across hundreds of frames. SAM 2’s promptable video object segmentation (Ravi et al., 2024) has demonstrated the architectural feasibility of video segmentation at ADE20K scale and makes video annotation using AI-assisted tools now tractable. Early planning for ADE20K-Video annotation is underway in the computer vision community, with proposed designs leveraging SAM 2 pre-segmentation for 80%+ of pixel labels and human correction for ambiguous regions.

• 3D scene parsing: As depth sensors (LiDAR, structured light, time-of-flight cameras) and neural multi-view reconstruction methods (NeRF, 3D Gaussian Splatting) become standard components of robotics platforms, AR headsets, and autonomous vehicle sensor suites, the research community needs ADE20K-style dense semantic annotations paired with 3D point cloud and mesh scene representations. A 3D-ADE20K benchmark labelling RGB-D scans of the same scene types currently covered by ADE20K’s 2D imagery — indoors at living spaces, kitchens, offices; outdoors at streets, parks, building facades — would enable models to learn unified 2D+3D Scene Understanding representations that ground pixel-level semantic categories in physically measured 3D space. Research at the University of Edinburgh’s VICO group, Oxford’s Active Vision Laboratory, and TU Munich’s Visual Computing group is already developing such multi-modal 2D+3D scene annotation pipelines that extend the ADE20K methodology to the depth dimension.

• Interactive and embodied segmentation: Rather than passive pixel labelling evaluated by a single forward-pass mIoU score, future ADE20K evaluations will assess models on iterative segmentation tasks where a human provides sparse prompt feedback — point clicks, bounding boxes, natural language corrections — and the model refines its prediction interactively across multiple rounds. SAM’s promptable interface (Kirillov et al., 2023) introduced this paradigm for interactive image segmentation, and future ADE20K benchmark variants will measure interactive segmentation efficiency (the number of human prompts required to reach a given mIoU quality threshold) alongside the conventional single-forward-pass mIoU metric, better capturing the practical utility of segmentation tools in annotation and deployment workflows.

• AI-in-the-loop domain extension: SAM pre-segmentation followed by expert human verification can reduce Data Annotation cost by 60–80% while maintaining quality comparable to fully manual annotation for well-defined category vocabularies. This makes it feasible to produce ADE20K-scale annotated corpora for specialised application domains — Medical Imaging scenes (radiology, pathology, endoscopy), industrial facilities (manufacturing floors, warehouses, power plants), agricultural environments (crop fields, orchards, greenhouses), marine and underwater environments — at costs and timescales previously prohibitive. Community-driven initiatives are already underway to produce domain-specific ADE20K extensions using AI-assisted annotation workflows, creating derivative benchmarks that inherit ADE20K’s evaluation protocol and category hierarchy conventions while targeting application-specific scene types and category vocabularies.

• Benchmark difficulty escalation and adversarial probing: The Foundation Models trend is likely to push ADE20K-150 mIoU scores above 70 within the 2026–2030 period, at which point the 150-category standard benchmark will lose its discriminative power as a research tool. The 847-category extended variant and adversarially curated hard subsets — images selected to feature unusual viewpoints, rare category instances, extreme lighting conditions, heavy occlusion, domain shift, and long-tail semantic concepts — will become the primary differentiating evaluation targets for distinguishing the capabilities of frontier models. Benchmark contamination research will simultaneously need to address whether models pre-trained on internet-scale image-text data have been trained on ADE20K images or their close visual derivatives, since contamination would inflate measured mIoU scores and require re-annotation of held-out test sets to restore measurement validity.

  Research & Literature

  1. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., & Torralba, A. (2017). Scene parsing through ADE20K dataset. CVPR 2017. https://openaccess.thecvf.com/content_cvpr_2017/html/Zhou_Scene_Parsing_Through_CVPR_2017_paper.html
  2. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., & Torralba, A. (2019). Semantic understanding of scenes through the ADE20K dataset. IJCV, 127(3), 302–321.
  3. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., & Torralba, A. (2018). Places: A 10 million image database for scene recognition. IEEE TPAMI, 40(6), 1452–1464.
  4. Xiao, J., Hays, J., Ehinger, K., Oliva, A., & Torralba, A. (2010). SUN database: Large-scale scene recognition. CVPR 2010.
  5. Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., … & Zitnick, C. L. (2014). Microsoft COCO: Common objects in context. ECCV 2014.
  6. Zhao, H., Shi, J., Qi, X., Wang, X., & Jia, J. (2017). Pyramid scene parsing network (PSPNet). CVPR 2017.
  7. Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018). Encoder-decoder with atrous separable convolution for semantic image segmentation (DeepLab v3+). ECCV 2018.
  8. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., … & Xiao, B. (2020). Deep high-resolution representation learning for visual recognition (HRNet). IEEE TPAMI, 43(10), 3349–3364.
  9. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., … & Zhang, L. (2021). Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers (SETR). CVPR 2021.
  10. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J. M., & Luo, P. (2021). SegFormer: Simple and efficient design for semantic segmentation with transformers. NeurIPS 2021.
  11. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., … & Guo, B. (2021). Swin Transformer: Hierarchical vision transformer using shifted windows. ICCV 2021.
  12. Cheng, B., Misra, I., Schwing, A. G., Kirillov, A., & Garg, R. (2022). Masked-attention mask transformer for universal image segmentation (Mask2Former). CVPR 2022.
  13. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., … & Wei, F. (2022). Image as a foreign language: BEiT pretraining for all vision tasks (BEiT-3). CVPR 2023.
  14. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., … & Qiao, Y. (2022). InternImage: Exploring large-scale vision foundation models with deformable convolutions. CVPR 2023.
  15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., … & Girshick, R. (2023). Segment anything. ICCV 2023.
  16. Ravi, N., Gabeur, V., Hu, Y.-T., Hu, R., Ryali, C., Ma, T., … & Feichtenhofer, C. (2024). SAM 2: Segment anything in images and videos. arXiv:2408.00714.
  17. Yu, X., Yin, X., Zhao, H., Shi, J., & Jia, J. (2023). FC-CLIP: Frozen CLIP for open-vocabulary segmentation. NeurIPS 2023.
  18. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., … & Marculescu, D. (2023). Open-vocabulary semantic segmentation with mask-adapted CLIP (OVSeg). CVPR 2023.
  19. Lüddecke, T., & Ecker, A. (2022). Image segmentation using text and image prompts (CLIPSeg). CVPR 2022.
  20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., … & Bojanowski, P. (2024). DINOv2: Learning robust visual features without supervision. TMLR 2024.
  21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., … & Houlsby, N. (2021). An image is worth 16×16 words: Transformers for image recognition at scale (ViT). ICLR 2021.
  22. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., & Girshick, R. (2022). Masked autoencoders are scalable vision learners (MAE). CVPR 2022.
  23. Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., & Xie, S. (2022). A ConvNet for the 2020s (ConvNeXt). CVPR 2022.
  24. Bau, D., Zhou, B., Khosla, A., Oliva, A., & Torralba, A. (2017). Network dissection: Quantifying interpretability of deep visual representations. CVPR 2017.
  25. Yao, H., Huang, R., Zhang, L., Wei, Y., Dong, L., Zhang, D., … & Wei, F. (2022). EVA: Exploring the limits of masked visual representation learning. CVPR 2023.
  26. Cheng, Z., Wang, Z., Lu, H., Liao, J., Fang, Y., & Ouyang, W. (2024). Foundation model assisted weakly supervised semantic segmentation. ResearchGate 2024. https://www.researchgate.net/publication/379713372
  27. Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., & Lu, H. (2019). Dual attention network for scene segmentation (DANet). CVPR 2019.

Provenance