Object Detection on PASCAL VOC
86.1mAPSemi-DETR (DINO)
Evaluation results
| Method | Links | ||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Semi-DETR (DINO)Category=End-to-End, Detector=DINO, Supervision=Semi-Supervised2023.07 | 86.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 65.2 | |
| DINO SSOD (Baseline)Category=End-to-End, Detector=DINO, Supervision=Semi-Supervised2023.07 | 84.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.2 | |
| Semi-DETR (Def-DETR)Category=End-to-End, Detector=Deformable DETR, Supervision=Semi-Supervised2023.07 | 83.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 57.2 | |
| Scale-aware AutoAugDetector=Faster R-CNN, Backbone=ResNet-502021.03 | 81.6 | 88.2 | 80.1 | 74.1 | 73.6 | 88.3 | 89.1 | 88.9 | 68.1 | 87.2 | 73.8 | 88.4 | 88.9 | 87.5 | 87.1 | 56.2 | 79 | 79.7 | 87.2 | 78.6 | 84.7 | — | |
| Faster-RCNN (D-efficient + cosine + smooth + mixup + distill)Refinement=+ distill w/ mixup, Backbone Top-1 Accuracy=79.292018.12 | 81.33 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Unbiased Teacher v2Category=One-Stage2023.07 | 81.29 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 56.87 | |
| DINO (Sup only)Category=End-to-End, Detector=DINO, Supervision=Supervised Only2023.07 | 81.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.6 | |
| Faster-RCNN (D-efficient + cosine + smooth + mixup w/o distill)Refinement=+ mixup w/o distill, Backbone Top-1 Accuracy=79.162018.12 | 81.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Faster-RCNN (D-efficient + cosine + smooth + distill w/o mixup)Refinement=+ distill w/o mixup, Backbone Top-1 Accuracy=78.672018.12 | 80.96 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QG-DMSGDModel=Faster R-CNN, Topology=Static, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 80.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Faster-RCNN (D-efficient + cosine + smooth)Refinement=+ smooth, Backbone Top-1 Accuracy=78.342018.12 | 80.71 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Vanilla DMSGDModel=Faster R-CNN, Topology=Static, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 80.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DSLCategory=One-Stage2023.07 | 80.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 56.8 | |
| Vanilla DMSGDModel=Faster R-CNN, Topology=One-Peer, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 80.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DMSGDModel=Faster R-CNN, Topology=One-Peer, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 80.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DMSGDModel=Faster R-CNN, Topology=Static, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 80.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QG-DMSGDModel=Faster R-CNN, Topology=One-Peer, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 80.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Parallel SGDModel=Faster R-CNN, Topology=Static, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 80.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Dense TeacherCategory=One-Stage2023.07 | 79.89 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 55.87 | |
| Faster-RCNN (D-efficient + cosine)Refinement=+ cosine, Backbone Top-1 Accuracy=77.912018.12 | 79.23 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QG-DMSGDModel=RetinaNet, Topology=Static, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 79.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Vanilla DMSGDModel=RetinaNet, Topology=One-Peer, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 79.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DMSGDModel=RetinaNet, Topology=Static, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 79.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QG-DMSGDModel=RetinaNet, Topology=One-Peer, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 79.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Parallel SGDModel=RetinaNet, Topology=Static, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 79 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Vanilla DMSGDModel=RetinaNet, Topology=Static, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 79 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DMSGDModel=RetinaNet, Topology=One-Peer, Backbone=ResNet-50, Batch Size=64, GPU Count=82021.10 | 79 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Def-DETR SSOD (Baseline)Category=End-to-End, Detector=Deformable DETR, Supervision=Semi-Supervised2023.07 | 78.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 53.4 | |
| Real-valuedFramework=Faster-RCNN, Backbone=ResNet-18, W/A=32/32, Memory Usage (MB)=112.88, OPs (×10^9)=96.40, Resolution=1000×6002026.05 | 78.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Faster R-CNN baselineDetector=Faster R-CNN, Backbone=ResNet-502021.03 | 78.6 | 80.8 | 79.3 | 72.3 | 67.2 | 87.4 | 88.5 | 88.6 | 62.6 | 86 | 71.2 | 88 | 88.9 | 80.6 | 79.9 | 52.6 | 78.7 | 74 | 86.2 | 78.3 | 80.9 | — | |
| HSAKDBackbone=ResNet-18, Teacher Model=ResNet-34, Framework=Faster-RCNN, Pre-training Dataset=ImageNet2021.07 | 78.45 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Faster-RCNN (D-efficient)Refinement=D-efficient, Backbone Top-1 Accuracy=77.162018.12 | 78.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSD512-RN50 (base model)FLOPs (G)=65.56, params (M)=21.97, FPS (BS=1)=68.24, FPS (BS=32)=103.482022.10 | 77.98 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSKDBackbone=ResNet-18, Teacher Model=ResNet-34, Framework=Faster-RCNN, Pre-training Dataset=ImageNet2021.07 | 77.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Faster-RCNN (B-standard)Refinement=B-standard, Backbone Top-1 Accuracy=76.142018.12 | 77.54 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSD512-RN50-HALPFLOPs (G)=15.38, params (M)=10.4, FPS (BS=1)=132.57, FPS (BS=32)=323.382022.10 | 77.42 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Unbiased TeacherCategory=Two-Stage2023.07 | 77.37 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 48.69 | |
| CRDBackbone=ResNet-18, Teacher Model=ResNet-34, Framework=Faster-RCNN, Pre-training Dataset=ImageNet2021.07 | 77.36 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RetinaNet-RN50FLOPs (G)=106.5, params (M)=36.5, FPS (BS=1)=36.922022.10 | 77.27 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KDBackbone=ResNet-18, Teacher Model=ResNet-34, Framework=Faster-RCNN, Pre-training Dataset=ImageNet2021.07 | 77.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SURGEFramework=Faster-RCNN, Backbone=ResNet-18, W/A=1/1, Memory Usage (MB)=16.61, OPs (×10^9)=18.49, Resolution=1000×6002026.05 | 77 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSD300-VGG16FLOPs (G)=31.44, params (M)=26.29, FPS (BS=1)=122.28, FPS (BS=32)=262.932022.10 | 76.72 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| IDa-DetFramework=Faster-RCNN, Backbone=ResNet-18, W/A=1/1, Memory Usage (MB)=16.61, OPs (×10^9)=18.49, Resolution=1000×6002026.05 | 76.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineBackbone=ResNet-18, Teacher Model=ResNet-34, Framework=Faster-RCNN, Pre-training Dataset=ImageNet2021.07 | 76.18 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSD512-RN50-slimFLOPs (G)=46.09, params (M)=16.33, FPS (BS=1)=76.49, FPS (BS=32)=114.82022.10 | 75.83 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSD300-RN50FLOPs (G)=16.23, params (M)=15.43, FPS (BS=1)=128.85, FPS (BS=32)=309.322022.10 | 75.69 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VICRegLBackbone=CNX-B, alpha=0.75, Evaluation Protocol=Fine-Tuned2022.10 | 75.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DINOBackbone=ViT-B, Evaluation Protocol=Fine-Tuned2022.10 | 74.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Def-DETR (Sup only)Category=End-to-End, Detector=Deformable DETR, Supervision=Supervised Only2023.07 | 74.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 46.2 | |
| DoReFa-NetFramework=Faster-RCNN, Backbone=ResNet-18, W/A=4/4, Memory Usage (MB)=21.59, OPs (×10^9)=27.15, Resolution=1000×6002026.05 | 73.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LWS-DetFramework=Faster-RCNN, Backbone=ResNet-18, W/A=1/1, Memory Usage (MB)=16.61, OPs (×10^9)=18.49, Resolution=1000×6002026.05 | 73.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VICRegBackbone=ViT-B, Evaluation Protocol=Fine-Tuned2022.10 | 71.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FasterRCNN-VGG16FLOPs (G)=91.23, params (M)=137.08, FPS (BS=1)=29.212022.10 | 70.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Moco-v3Backbone=ViT-B, Evaluation Protocol=Fine-Tuned2022.10 | 69.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ReActNetFramework=Faster-RCNN, Backbone=ResNet-18, W/A=1/1, Memory Usage (MB)=16.61, OPs (×10^9)=18.49, Resolution=1000×6002026.05 | 69.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Siamese DETRArchitecture=DN-DETR, Initialization=Siamese DETR pre-trained2023.03 | 63.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Siamese DETRArchitecture=DAB-DETR, Initialization=Siamese DETR pre-trained2023.03 | 62.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VICRegLBackbone=R50, alpha=0.75, Evaluation Protocol=Fine-Tuned2022.10 | 59.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VICRegLBackbone=CNX-B, alpha=0.75, Evaluation Protocol=Linear Detection2022.10 | 59 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DN-DETRInitialization=from scratch2023.03 | 58.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseCLBackbone=R50, Evaluation Protocol=Fine-Tuned2022.10 | 58.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DINOBackbone=ViT-B, Evaluation Protocol=Linear Detection2022.10 | 58.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DAB-DETRInitialization=from scratch2023.03 | 57.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VICRegBackbone=R50, Evaluation Protocol=Fine-Tuned2022.10 | 57.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCo v2Backbone=R50, Evaluation Protocol=Fine-Tuned2022.10 | 57 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ImageNet labelsevaluation protocol=ALL2018.07 | 56.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DeepClusterevaluation protocol=ALL2018.07 | 55.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VICRegBackbone=ViT-B, Evaluation Protocol=Linear Detection2022.10 | 55.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Grid R-CNNbackbone=ResNet-50, framework=FPN based2018.11 | 55.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Moco-v3Backbone=ViT-B, Evaluation Protocol=Linear Detection2022.10 | 54.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Noroozi and Favaroevaluation protocol=ALL2018.07 | 53.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Pathak et al.evaluation protocol=ALL2018.07 | 52.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FPNbackbone=ResNet-502018.11 | 51.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Noroozi et al.evaluation protocol=ALL2018.07 | 51.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DeepClusterevaluation protocol=FC6-82018.07 | 51.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Doersch et al.evaluation protocol=ALL, data-dependent initialization=true2018.07 | 51.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Bojanowski and Joulinevaluation protocol=ALL, data-dependent initialization=true2018.07 | 49.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GrayTarget Detector=SSD2026.04 | 49.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GrayTarget Detector=YOLOv52026.04 | 48.03 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Random-sobelevaluation protocol=ALL2018.07 | 47.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Random NoiseTarget Detector=YOLOv52026.04 | 47.39 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Wang and Guptaevaluation protocol=ALL, data-dependent initialization=true2018.07 | 47.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Random NoiseTarget Detector=SSD2026.04 | 47.16 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Donahue et al.evaluation protocol=ALL, data-dependent initialization=true2018.07 | 46.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Zhang et al.evaluation protocol=ALL, data-dependent initialization=true2018.07 | 46.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Zhang et al.evaluation protocol=ALL, data-dependent initialization=true2018.07 | 46.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| R-FCNbackbone=ResNet-502018.11 | 45.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VICRegLBackbone=R50, alpha=0.75, Evaluation Protocol=Linear Detection2022.10 | 45.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Random NoiseTarget Detector=Avg.2026.04 | 45.25 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Random-rgbevaluation protocol=ALL2018.07 | 44.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Pathak et al.evaluation protocol=ALL2018.07 | 44.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GrayTarget Detector=Avg.2026.04 | 44.39 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WhiteTarget Detector=SSD2026.04 | 43.84 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Zhang et al.evaluation protocol=FC6-8, data-dependent initialization=true, produced by=current authors2018.07 | 43.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseCLBackbone=R50, Evaluation Protocol=Linear Detection2022.10 | 43 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VICRegBackbone=R50, Evaluation Protocol=Linear Detection2022.10 | 41.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCo v2Backbone=R50, Evaluation Protocol=Linear Detection2022.10 | 41.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Random NoiseTarget Detector=Faster R-CNN2026.04 | 41.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WhiteTarget Detector=YOLOv52026.04 | 40.46 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WhiteTarget Detector=Avg.2026.04 | 38.97 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |