Bounding Box Regression

Bounding Box Regression is the supervised learning task within object detection of predicting the precise coordinates of axis-aligned rectangular boxes that tightly enclose detected objects. A model outputs four continuous values — typically centre x, centre y, width, and height relative to an anchor — and is trained using smooth-L1 or IoU-based loss functions that penalise deviation from ground-truth boxes. It is jointly trained with an object classification head and combined with Non Maximum Suppression to produce the final detections. Accurate regression is critical for downstream tasks such as instance segmentation and 3-D pose estimation that consume the predicted boxes.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:hasPart ai:NonMaximumSuppression))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:hasPart ai:FeaturePyramidNetwork))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:hasPart ai:AnchorBox))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:hasPart ai:IoULoss))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:hasPart ai:HeatmapRegression))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:hasPart ai:RoIPooling))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:hasPart ai:RegressionHead))

Dependency Relationships

SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:dependsOn ai:ConvolutionalNeuralNetwork))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:dependsOn ai:Backpropagation))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:dependsOn ai:TrainingData))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:dependsOn ai:GradientDescent))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:requires ai:LossFunction))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:requires ai:BenchmarkDataset))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:requires ai:ComputerVision))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:uses ai:TransformerArchitecture))

Capability Relationships

SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:enables ai:ObjectDetectionAndTracking))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:enables ai:InstanceSegmentation))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:enables ai:PoseEstimation))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:enables ai:AutonomousDriving))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:enables ai:SceneUnderstanding))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:supports ai:RealTimeAIInference))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:supports ai:AugmentedReality))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:supports ai:MedicalImageAnalysis))

Implementation Relationships

SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:implements ai:ObjectDetection))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:implements ai:Localisation))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:partOf ai:ObjectDetection))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:partOf ai:ComputerVisionPipeline))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:uses ai:FeaturePyramidNetwork))

Reduction Relationships

SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:reducesTo ai:Regression))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:reducesTo ai:CoordinatePrediction))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:contrastsWith ai:HeatmapRegression))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:contrastsWith ai:SemanticSegmentation))
SubClassOf(ai:BoundingBoxRegression
  ObjectSomeValuesFrom(ai:contrastsWith ai:RegressionTesting))

About

Bounding box regression emerged as a formalised subtask with the R-CNN family of detectors introduced by Ross Girshick and colleagues at UC Berkeley in 2014. The original R-CNN paper applied a post-hoc linear regression correction to region proposals generated by selective search, demonstrating that even a simple linear transformation of CNN features could substantially improve localisation accuracy beyond what classification-oriented CNNs produced by default. Fast R-CNN (Girshick, 2015) folded both the regression and classification into a single end-to-end trained network using RoI pooling, and introduced the smooth-L1 loss function — a hybrid loss that transitions from L2 (squared error) for small residuals to L1 (absolute error) for large residuals, providing outlier robustness without sacrificing gradient magnitude near zero. Faster R-CNN (Ren et al., 2015) replaced selective search with a Region Proposal Network (RPN) that itself performs bounding box regression to propose candidate regions, making the entire detection pipeline end-to-end differentiable for the first time.

The subsequent evolution of bounding box regression has been driven by the recognition that coordinate-based losses — smooth-L1, L2 — do not directly optimise the evaluation metric (IoU) and may produce perverse gradients when predicted and ground-truth boxes do not overlap. IoU-based losses address this by computing the geometric overlap directly: GIoU (Rezatofighi et al., 2019) adds a penalty for the smallest enclosing box to provide gradients when boxes do not overlap; DIoU (Zheng et al., 2020) additionally penalises the distance between box centres; CIoU further adds an aspect-ratio consistency term; and subsequent work has proposed SIoU, EIoU, WIoU, and Powerful-IoU (Neural Networks, 2024), each targeting specific failure modes. As of 2025–2026, CIoU has been adopted as the default in YOLOv8 and YOLO11, while the YOLO26 release (Ultralytics, October 2025) introduced Distribution Focal Loss (DFL) removal in favour of cleaner coordinate regression combined with the Progressive Loss (ProgLoss) and Spatial Token Alignment Loss (STAL) for improved small-object detection accuracy.

Anchor-free detectors constitute a parallel development strand. FCOS (Tian et al., 2019) predicts the distance from each foreground pixel to the four sides of the enclosing bounding box, eliminating the anchor hyperparameter tuning burden. CenterNet (Zhou et al., 2019) and CenterPoint (Yin et al., 2021) predict object centres as heatmap peaks and regress size offsets from those centres, achieving state-of-the-art performance on 3D object detection benchmarks for autonomous driving. DETR (Carion et al., 2020) and its successors (Deformable DETR, DAB-DETR, DN-DETR, DINO-DETR) eliminate both anchors and NMS by framing detection as a direct set prediction problem via Hungarian matching: each query predicts one object, and the box regression loss is computed using a combination of L1 and GIoU losses on matched pairs. NAN-DETR (2024) improved convergence via multi-anchor strategies and CIoU integration.

Components / Architecture

The bounding box regression pipeline in a modern detector contains the following components:

• Backbone Network: A Convolutional Neural Network (ResNet, EfficientNet, CSP-DarkNet) or Transformer Architecture (ViT, Swin) that extracts hierarchical feature representations from the input image; modern detectors increasingly use hybrid CNN–Transformer backbones

• Feature Pyramid Network (FPN): A top-down pathway with lateral connections that constructs multi-scale feature maps at strides 8, 16, 32, 64, and 128 pixels, enabling simultaneous regression of small and large objects from appropriately scaled features

• Anchor Box (anchor-based only): Predefined rectangular templates of multiple scales and aspect ratios tiled at each feature map location; the regression head predicts offsets (Δx, Δy, Δw, Δh) from these templates, normalised to reduce numerical scale variation; modern YOLO variants use learned anchor clusters from k-means on dataset bounding boxes

• Regression Head: A small convolutional (or MLP, in transformers) subnetwork attached to each FPN level that outputs four continuous values per spatial location (or per query, in DETR). YOLO architectures share the backbone and neck but fork into parallel classification and regression branches at the detection head

• Loss Function: The training objective for box regression; evolution: L2 → smooth-L1 (Fast R-CNN, 2015) → GIoU (2019) → DIoU/CIoU (2020) → EIoU/SIoU/WIoU (2021–2022) → Powerful-IoU (Neural Networks, 2024) → ProgLoss+STAL (YOLO26, 2025). IoU-family losses directly optimise the evaluation metric and provide informative gradients even when boxes partially overlap

• Anchor Assignment: The rule that assigns ground-truth boxes to anchors (or queries in DETR) for training: IoU-threshold assignment (anchor-based), ATSS (adaptive training sample selection), SimOTA (YOLOv6/7), or Hungarian matching (DETR family)

• Non Maximum Suppression: Post-processing (anchor-based and one-stage detectors only) that suppresses duplicate high-IoU detections by retaining only the highest-scoring box among overlapping candidates; variants include Soft-NMS, weighted-NMS, and class-agnostic NMS. DETR eliminates NMS entirely through set-prediction matching

• RoI Pooling / RoI Align: In two-stage detectors (Faster R-CNN, Mask R-CNN), a fixed-size feature is extracted from each proposal region using bilinear interpolation (RoI Align) and fed to a second-stage regression head for refinement; RoI Align (He et al., 2017) replaced quantised RoI Pooling to eliminate spatial misalignment

  Use Cases / Major Families

  Two-Stage Detectors (R-CNN Family) R-CNN (2014), Fast R-CNN (2015), Faster R-CNN (2015), and Mask R-CNN (2017) established the two-stage paradigm: a region proposal stage generates candidate boxes, which are refined by a second-stage regression and classification head. Mask R-CNN added an instance segmentation branch alongside the box head, enabling simultaneous Instance Segmentation. Cascade R-CNN (Cai & Vasconcelos, 2018) runs multiple successive regression stages at increasing IoU thresholds for high-precision localisation. These detectors dominate tasks requiring precision: medical image analysis (tumour localisation in CT/MRI), document layout analysis, and scientific image annotation pipelines.

  Single-Stage Detectors (YOLO Family) YOLO (You Only Look Once, Redmon et al., 2016) introduced the paradigm of predicting boxes directly from a single convolutional pass over the image, enabling real-time inference. YOLOv2 (2017) added anchor boxes and multi-scale training; YOLOv3 (2018) adopted FPN; YOLOv4 (2020) added mosaic augmentation and CSP; YOLOv5 (Ultralytics, 2020) was the first widely adopted open-source PyTorch YOLO. YOLOv8 (Ultralytics, 2023) adopted an anchor-free head and CIoU loss. YOLO11 (Ultralytics, 2024) refined the architecture. YOLO26 (Ultralytics, October 2025) natively unifies detection, instance segmentation, classification, pose estimation, and oriented bounding box (OBB) detection in a single model, achieving state-of-the-art performance on COCO while removing DFL in favour of ProgLoss+STAL. The YOLO family dominates real-time applications: autonomous vehicles, edge AI, video surveillance, and industrial inspection.

  Single-Stage Detectors (SSD and RetinaNet) SSD (Liu et al., 2016) uses multi-scale convolutional feature maps with default boxes at multiple aspect ratios, enabling detection without a proposal stage. RetinaNet (Lin et al., 2017) introduced the Focal Loss to address class imbalance in dense detection, achieving accuracy competitive with two-stage detectors at single-stage speed. EfficientDet (Tan et al., 2020) combined EfficientNet backbones with weighted BiFPN for a scalable family of detectors.

  Anchor-Free Detectors (FCOS, CenterNet Family) FCOS (Tian et al., 2019) predicts left/top/right/bottom distances from foreground pixels, with centerness scores to suppress low-quality peripheral predictions. CenterNet (Objects as Points, 2019) predicts keypoint heatmaps for object centres and regresses width/height offsets, providing a clean, anchor-free formulation. CenterPoint (2021) extended this to LiDAR-based 3D bounding box regression for autonomous driving, predicting box parameters in bird’s-eye-view feature maps.

  Transformer-Based Detectors (DETR Family) DETR (Detection Transformer, Carion et al., 2020) eliminated anchor design and NMS, replacing them with a fixed set of learned queries matched to ground-truth boxes via the Hungarian algorithm. Deformable DETR (Zhu et al., 2021) addressed convergence speed with deformable attention. DAB-DETR (Liu et al., 2022) and DN-DETR (Li et al., 2022) improved training stability. DINO (Zhang et al., 2022) achieved SOTA on COCO and spawned the current dominant family. MonoDINO-DETR (2025) extended DETR principles to monocular 3D object detection using vision foundation model features. A comprehensive 2025 survey (MDPI Sensors, PMC12252279) reviewed the full DETR family and its applications.

  Specialised Applications

• Medical Imaging: Faster R-CNN and Mask R-CNN variants are widely deployed for tumour localisation in CT, MRI, and histopathology images; YOLO variants enable real-time surgical instrument detection in laparoscopy

• Autonomous Driving: CenterPoint (LiDAR), BEVFusion (camera+LiDAR), and BEV-DETR families predict 3D bounding boxes in bird’s-eye-view coordinates for vehicles, pedestrians, and cyclists; evaluated on nuScenes and Waymo Open Dataset benchmarks

• Aerial and Satellite Imagery: Oriented bounding box (OBB) regression predicts rotated rectangles for arbitrarily oriented objects; YOLO26 and RT-DETR support OBB natively

• Industrial Quality Inspection: YOLO-based detectors deployed on edge GPUs (Jetson Orin) detect manufacturing defects at production-line speeds

• Augmented Reality: Bounding box regression provides the spatial anchor for virtual object placement; Apple Vision Pro, Meta Quest 3, and HoloLens 2 use on-device detection models

  Academic Context

  Bounding box regression as a formalised task emerged from the Deformable Part Models (DPM) era. Felzenszwalb et al.’s DPM (CVPR 2008, IEEE TPAMI 2010) used bounding box regression within a latent SVM framework to predict object extents. Uijlings et al.’s selective search (IJCV 2013) provided region proposals consumed by early CNN detectors.

  The deep learning era was inaugurated by Girshick et al.’s R-CNN (CVPR 2014), which demonstrated that CNN features with post-hoc box regression dramatically outperformed DPM on PASCAL VOC. Fast R-CNN (Girshick, ICCV 2015) introduced smooth-L1 loss and end-to-end training. Faster R-CNN (Ren, He, Girshick, Sun, NeurIPS 2015) achieved real-time proposal generation with the RPN. He et al.’s Mask R-CNN (ICCV 2017) unified detection, instance segmentation, and keypoint estimation in a single framework, winning the COCO 2017 detection and segmentation challenges.

  The YOLO lineage (Redmon et al., CVPR 2016; Redmon & Farhadi, CVPR 2017, 2018) demonstrated that single-stage direct regression could approach two-stage accuracy at an order-of-magnitude speedup, enabling real-time deployment. RetinaNet (Lin et al., ICCV 2017) and the Focal Loss resolved the class-imbalance problem that had limited single-stage accuracy.

  IoU-based losses were pioneered by Yu et al. (IoU Loss, 2016), followed by GIoU (Rezatofighi et al., CVPR 2019), DIoU/CIoU (Zheng et al., AAAI 2020), EIoU (Zhang et al., 2022), SIoU (Gevorgyan, 2022), and Powerful-IoU (Neural Networks, 2024). Each generation addressed residual failure modes: non-overlapping box gradients, centre distance convergence, aspect ratio consistency, and outlier-sensitive focusing.

  The anchor-free paradigm was systematised by FCOS (Tian et al., ICCV 2019) and CenterNet (Zhou et al., 2019). DETR (Carion et al., ECCV 2020) introduced the transformer paradigm; the subsequent lineage (Deformable DETR 2021, DAB-DETR 2022, DN-DETR 2022, DINO 2022, RT-DETR 2023) has converged on deformable attention with contrastive denoising as the dominant approach. As of 2026, DINO-based detectors hold COCO instance segmentation and panoptic records.

  COCO (Lin et al., ECCV 2014) became the canonical benchmark, with the AP@0.5:0.95 (COCO mAP) metric replacing PASCAL VOC AP@0.5. Waymo Open Dataset and nuScenes establish 3D detection benchmarks for autonomous driving evaluation.

  Current Landscape (2026)

  As of mid-2026, bounding box regression is a mature but actively evolving subfield:

• YOLO26 (Ultralytics, October 2025): The latest YOLO generation is the first to natively unify five tasks (detection, instance segmentation, classification, pose estimation, oriented bounding box detection) in a single architecture. It replaces DFL with cleaner coordinate regression, introduces ProgLoss for training stability, and STAL for small-object accuracy. YOLO26 instance segmentation extends the detection head with a prototype mask branch for pixel-level masks in real time

• RT-DETR v2 (2025): Baidu’s real-time DETR family achieves competitive accuracy to YOLO26 with transformer efficiency improvements via multi-scale hybrid attention, enabling deployment on automotive-grade hardware

• Vision Foundation Models: SAM 2 (Meta, 2024) and Grounding DINO enable zero-shot and open-vocabulary bounding box prediction from text or click prompts, extending beyond closed-vocabulary detection to any object; YOLO-SAM integration pipelines (Scientific Reports, 2025) combine real-time YOLO detection with SAM segmentation for joint detection and segmentation

• Small Object Detection: SOD-YOLOv8 (2024) and PCPE-YOLO (2025, Nature Scientific Reports) address the longstanding challenge of small object detection in traffic scenes with new neck architectures and loss formulations; Corner-Point and Foreground-Area IoU loss (2023, PMC) specifically targets small-object localisation

• 3D Bounding Box Regression: BEVFusion, BEV-DETR, and MonoDINO-DETR advance monocular and fusion-based 3D detection, critical for autonomous driving validation on the Waymo and nuScenes leaderboards

• Edge Deployment: YOLO26-nano and similar compressed variants run on Jetson Orin NX (16 TOPS) and Apple Neural Engine (38 TOPS on M4 chip) at 30+ FPS for on-device deployment in autonomous vehicles, drones, and AR headsets

• Oriented Bounding Box (OBB): Satellite imagery and drone aerial analysis require rotated bounding boxes; YOLO26, RT-DETR, and dedicated detectors like Oriented R-CNN provide OBB regression natively, with evaluation on DOTA and HRSC2016 benchmarks

  UK Context

  The United Kingdom has several active research groups and industrial programmes in bounding box regression and object detection. At the University of Oxford, the Visual Geometry Group (VGG), which originated VGGNet, continues influential work in detection, video understanding, and visual grounding. Oxford’s Active Vision Group contributed detection work for robotic manipulation. At the University of Cambridge, the Computer Laboratory’s computer vision group works on detection for medical imaging and autonomous vehicles.

  Imperial College London’s Dyson Robotics Lab integrates real-time object detection with SLAM (simultaneous localisation and mapping) for robotic scene understanding, with bounding box regression serving as the localisation signal for grasping. University College London (UCL) has contributed work on few-shot and weakly supervised detection relevant to reducing annotation cost.

  In Northern England, the University of Manchester’s School of Computer Science has research in medical image detection (histopathology, radiology) that relies heavily on bounding box regression for tumour and lesion localisation. Leeds Institute for Data Analytics (LIDA) works on aerial image object detection for environmental monitoring, including flood mapping and agricultural analysis using satellite imagery. The University of Sheffield’s Vision, Speech and Signal Processing group has applied detection models to industrial quality assurance in the Sheffield Advanced Manufacturing Research Centre (AMRC) context.

  Newcastle University’s Open Lab and School of Computing have worked on detection for assistive technologies — object localisation for blind and visually impaired users via wearable cameras — a UK-funded programme under EPSRC. Edinburgh’s Informatics group has contributed to transformer-based detection research.

  Industry presence includes Wayve (London), the autonomous driving company that uses camera-first detection pipelines for end-to-end driving; Five AI (Bristol), acquired by Bosch, which developed UK-specific autonomous driving datasets; and Oxbotica (Oxford), which built LiDAR and camera fusion detection systems for logistics and mining vehicles. ARM Holdings (Cambridge) designs the neural processing units embedded in automotive SoCs that execute bounding box regression at the edge. DeepMind (London) has contributed to detection-adjacent work in video understanding and Gemini Vision capabilities.

  Future Directions (2026–2030)

• End-to-end 3D detection from cameras: Eliminating LiDAR dependence by predicting full 3D bounding boxes (x, y, z, l, w, h, heading) from monocular or multi-camera systems using depth foundation models and occupancy prediction; TESLA’s pure-vision Full Self-Driving architecture drives this direction commercially

• Open-vocabulary and grounded detection: Extending bounding box regression beyond closed-category vocabularies to any natural language description using vision-language models (Grounding DINO, GLIP, OWL-ViT), enabling detection of arbitrary objects at test time without retraining

• Neuromorphic and event-camera detection: Detecting and regressing boxes from event camera streams (microsecond temporal resolution, sparse spatial activation) using spiking neural networks for ultra-low-latency, low-power autonomous systems

• Panoptic foundation models: Unified architectures that simultaneously perform detection, instance segmentation, panoptic segmentation, and semantic segmentation in a single forward pass, replacing task-specific heads with a general decoder

• 4D bounding boxes: Predicting spatiotemporal bounding tubes in video — x, y, w, h, start-frame, end-frame — for activity localisation and video object segmentation, extending the 2D box paradigm to the temporal dimension

• Federated and privacy-preserving detection: Training bounding box regression models on distributed medical or surveillance data without centralising patient images, using federated learning and differential privacy; critical for UK NHS imaging data governance

• Regulatory integration: EU AI Act Annex III classifies safety-critical CV systems (autonomous vehicle perception, medical devices) as high-risk; bounding box regression models in these domains will require conformity assessments, robustness evaluations, and ongoing monitoring post-deployment

  Benchmark Datasets

• COCO (MS-COCO): 118,000 training images, 80 object categories, 5 annotations per image including bounding boxes, instance masks, and captions; primary benchmark since 2014; metric is COCO mAP (AP@0.5:0.95)

• Pascal VOC: Earlier benchmark (2005–2012), 20 categories, metric is AP@0.5; widely used for reproducibility comparison

• Open Images v7: 9 million images, 600 categories, bounding box and relationship annotations; larger than COCO

• LVIS: Long-tail distribution benchmark with 1,203 categories for evaluating detection of rare objects

• Waymo Open Dataset: 1,950 scenes, LiDAR + camera, for autonomous driving 2D and 3D detection

• nuScenes: 1,000 driving scenes, 6 cameras + LiDAR + radar, for 3D detection and tracking

• DOTA: Aerial and satellite imagery with oriented bounding boxes for remote sensing detection

• VisDrone2019: Drone-captured imagery for small-object detection evaluation; used in PCPE-YOLO and SOD-YOLOv8 evaluations

  Research & Literature

  1. Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich feature hierarchies for accurate object detection and semantic segmentation. CVPR 2014, 580–587. https://doi.org/10.1109/CVPR.2014.81
  2. Girshick, R. (2015). Fast R-CNN. ICCV 2015, 1440–1448. https://doi.org/10.1109/ICCV.2015.169
  3. Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-CNN: Towards real-time object detection with region proposal networks. NeurIPS 2015, 91–99.
  4. He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN. ICCV 2017, 2961–2969. https://doi.org/10.1109/ICCV.2017.322
  5. Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. CVPR 2016, 779–788. https://doi.org/10.1109/CVPR.2016.91
  6. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y., & Berg, A. C. (2016). SSD: Single shot multibox detector. ECCV 2016, 21–37.
  7. Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017). Focal loss for dense object detection. ICCV 2017, 2980–2988. https://doi.org/10.1109/ICCV.2017.324
  8. Tian, Z., Shen, C., Chen, H., & He, T. (2019). FCOS: Fully convolutional one-stage object detection. ICCV 2019, 9627–9636. https://doi.org/10.1109/ICCV.2019.00972
  9. Zhou, X., Wang, D., & Krähenbühl, P. (2019). Objects as points. arXiv:1904.07850.
  10. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020). End-to-end object detection with transformers. ECCV 2020, 213–229. https://doi.org/10.1007/978-3-030-58452-8_13
  11. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., & Savarese, S. (2019). Generalized intersection over union: A metric and a loss for bounding box regression. CVPR 2019, 658–666. https://doi.org/10.1109/CVPR.2019.00075
  12. Zheng, Z., Wang, P., Liu, W., Li, J., Ye, R., & Ren, D. (2020). Distance-IoU loss: Faster and better learning for bounding box regression. AAAI 2020, 12993–13000.
  13. 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, 740–755.
  14. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., & Dai, J. (2021). Deformable DETR: Deformable transformers for end-to-end object detection. ICLR 2021.
  15. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., & Shum, H.-Y. (2022). DINO: DETR with improved denoising anchor boxes for end-to-end object detection. ICLR 2023. https://arxiv.org/abs/2203.03605
  16. Yin, T., Zhou, X., & Krähenbühl, P. (2021). Center-based 3D object detection and tracking. CVPR 2021, 11784–11793.
  17. Wang, C.-Y., Bochkovskiy, A., & Liao, H.-Y. M. (2023). YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. CVPR 2023, 7464–7475.
  18. Ultralytics. (2025). YOLO26: Unified multi-task real-time object detection. https://playground.roboflow.com/models/ultralytics/yolo26
  19. Ultralytics. (2025). YOLO26 instance segmentation. LearnOpenCV. https://learnopencv.com/yolo26-instance-segmentation-pixel-perfect-ai-at-real-time-speed/
  20. PMC. (2025). A review of DEtection TRansformer: From basic architecture to advanced developments. MDPI Sensors, 25(13), 3952. https://pmc.ncbi.nlm.nih.gov/articles/PMC12252279/
  21. Zhang, B., et al. (2024). Powerful-IoU: More straightforward and faster bounding box regression loss with a nonmonotonic focusing mechanism. Neural Networks, 170. https://doi.org/10.1016/j.neunet.2023.11.041
  22. ResearchGate. (2025). Enhancing bounding box regression for object detection: Dimensional angle precision IoU-loss. https://www.researchgate.net/publication/391528309
  23. Nature Scientific Reports. (2025). PCPE-YOLO with a lightweight dynamically reconfigurable backbone for small object detection. https://www.nature.com/articles/s41598-025-15975-w
  24. Nature Scientific Reports. (2025). YOLO-SAM: An end-to-end framework for efficient real-time object detection and segmentation. https://www.nature.com/articles/s41598-025-24576-6
  25. arxiv. (2025). SOD-YOLOv8: Enhancing YOLOv8 for small object detection in traffic scenes. https://arxiv.org/pdf/2408.04786
  26. arxiv. (2025). MonoDINO-DETR: Depth-enhanced monocular 3D object detection using a vision foundation model. https://arxiv.org/pdf/2502.00315
  27. PMC. (2023). Corner-point and foreground-area IoU loss: Better localisation of small objects in bounding box regression. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10223589/
  28. NAN-DETR. (2024). Noising multi-anchor makes DETR better for object detection. PMC. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11513373/

  Key Terminology

• Bounding Box: An axis-aligned rectangle defined by (x_min, y_min, x_max, y_max) or equivalently (cx, cy, w, h) that encloses a detected object

• Anchor Box: A predefined rectangle of fixed scale and aspect ratio at each feature map location, used as a reference from which regression targets are computed as offsets

• Intersection over Union (IoU): The ratio of the intersection area to the union area of two boxes; the primary evaluation metric for bounding box quality; higher is better

• Smooth-L1 Loss: A piecewise loss function that transitions from L2 for |error| < 1 to L1 for |error| ≥ 1, combining smoothness near zero with outlier robustness

• GIoU Loss: Generalised IoU loss that adds a penalty for the enclosing box to provide non-zero gradients when boxes do not overlap

• CIoU Loss: Complete IoU loss that adds centre-distance and aspect-ratio consistency terms to DIoU, improving convergence

• NMS (Non-Maximum Suppression): A greedy post-processing step that iteratively selects the highest-scoring box and suppresses overlapping candidates above an IoU threshold

• FPN (Feature Pyramid Network): A top-down architecture that combines high-resolution low-level features with high-level semantic features via lateral connections for multi-scale detection

• Anchor-Free Detection: Paradigms (FCOS, CenterNet) that predict boxes directly from feature pixels without predefined anchor templates

• mAP (Mean Average Precision): The primary detection metric; averaged over categories and IoU thresholds (COCO uses 0.5:0.05:0.95)

• DFL (Distribution Focal Loss): A regression loss that models box coordinates as discrete distributions rather than point estimates, improving calibration; removed in YOLO26 in favour of ProgLoss

Provenance