| Rank | Model |
mask AP | AP50 | AP75 | APL | APM | APS | GFLOPs | Params (M) |
Extra Training Data |
Paper | Code | Year |
| 1 |
Co-DETR |
56.6 | 79.7 | 62.8 | 74.6 | 59.7 | 38.9 | – | – |
✓ |
DETRs with Collaborative Hybrid Assignments Training
|
open-mmlab/mmdetection · siyuanliii/masa · sense-x/co-detr
· +3 |
2022 |
| 2 |
ViT-CoMer-L (Mask RCNN, DINOv2) |
55.9 | – | – | – | – | – | – | – |
|
ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions
|
Traffic-X/ViT-CoMer · chenller/mmseg-extension |
2024 |
| 3 |
InternImage-H |
55.4 | 80.1 | 61.5 | 74.4 | 58.4 | 37.9 | – | – |
✓ |
InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions
|
opengvlab/internimage · OpenGVLab/M3I-Pretraining · chenller/mmseg-extension |
2022 |
| 4 |
EVA |
55.0 | 79.4 | 60.9 | 72.0 | 58.4 | 37.6 | – | – |
✓ |
EVA: Exploring the Limits of Masked Visual Representation Learning at Scale
|
rwightman/pytorch-image-models · open-mmlab/mmselfsup · baaivision/eva
· +3 |
2022 |
| 5 |
Mask Frozen-DETR |
54.9 | 78.9 | 60.8 | 72.9 | 58.4 | 37.2 | – | – |
✓ |
Mask Frozen-DETR: High Quality Instance Segmentation with One GPU
|
|
2023 |
| 6 |
MasK DINO (SwinL, multi-scale) |
54.5 | – | – | – | – | – | – | – |
✓ |
Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation
|
PaddlePaddle/PaddleDetection · IDEACVR/DINO · idea-research/maskdino
· +7 |
2022 |
| 7 |
ViT-Adapter-L (HTC++, BEiTv2, O365, multi-scale) |
54.2 | – | – | – | – | – | – | – |
✓ |
Vision Transformer Adapter for Dense Predictions
|
czczup/vit-adapter · chenller/mmseg-extension |
2022 |
| 7 |
GLEE-Pro |
54.2 | – | – | – | – | – | – | – |
✓ |
General Object Foundation Model for Images and Videos at Scale
|
FoundationVision/GLEE |
2023 |
| 9 |
SwinV2-G (HTC++) |
53.7 | – | – | – | – | – | – | – |
✓ |
Swin Transformer V2: Scaling Up Capacity and Resolution
|
rwightman/pytorch-image-models · microsoft/Swin-Transformer · PaddlePaddle/PaddleDetection
· +20 |
2021 |
| 10 |
ViTDet, ViT-H Cascade (multiscale) |
53.1 | – | – | – | – | – | – | – |
|
Exploring Plain Vision Transformer Backbones for Object Detection
|
facebookresearch/detectron2 · PaddlePaddle/PaddleDetection · alibaba/EasyCV
· +8 |
2022 |
| 11 |
GLEE-Plus |
53.0 | – | – | – | – | – | – | – |
✓ |
General Object Foundation Model for Images and Videos at Scale
|
FoundationVision/GLEE |
2023 |
| 12 |
Mask DINO (SwinL) |
52.6 | – | – | – | – | – | – | – |
|
Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation
|
PaddlePaddle/PaddleDetection · IDEACVR/DINO · idea-research/maskdino
· +7 |
2022 |
| 13 |
Soft Teacher + Swin-L(HTC++, multi-scale) |
52.5 | – | – | – | – | – | – | – |
✓ |
End-to-End Semi-Supervised Object Detection with Soft Teacher
|
microsoft/SoftTeacher · amazon-science/bigdetection · amazon-research/bigdetection
· +5 |
2021 |
| 13 |
ViT-Adapter-L (HTC++, BEiTv2 pretrain, multi-scale) |
52.5 | – | – | – | – | – | – | – |
|
Vision Transformer Adapter for Dense Predictions
|
czczup/vit-adapter · chenller/mmseg-extension |
2022 |
| 15 |
ViT-Adapter-L (HTC++, BEiT pretrain, multi-scale) |
52.2 | – | – | – | – | – | – | – |
|
Vision Transformer Adapter for Dense Predictions
|
czczup/vit-adapter · chenller/mmseg-extension |
2022 |
| 16 |
ViTDet, ViT-H Cascade |
52 | – | – | – | – | – | – | – |
|
Exploring Plain Vision Transformer Backbones for Object Detection
|
facebookresearch/detectron2 · PaddlePaddle/PaddleDetection · alibaba/EasyCV
· +8 |
2022 |
| 17 |
Soft Teacher + Swin-L(HTC++, single-scale) |
51.9 | – | – | – | – | – | – | – |
✓ |
End-to-End Semi-Supervised Object Detection with Soft Teacher
|
microsoft/SoftTeacher · amazon-science/bigdetection · amazon-research/bigdetection
· +5 |
2021 |
| 18 |
CBNetV2 (Dual-Swin-L HTC, multi-scale) |
51.8 | – | – | – | – | – | – | – |
|
CBNet: A Composite Backbone Network Architecture for Object Detection
|
PaddlePaddle/PaddleDetection · shinya7y/UniverseNet · VDIGPKU/CBNetV2
· +1 |
2021 |
| 19 |
Frozen Backbone, SwinV2-G-ext22K (HTC) |
51.6 | – | – | – | – | – | – | – |
|
Could Giant Pretrained Image Models Extract Universal Representations?
|
|
2022 |
| 20 |
CBNetV2 (Dual-Swin-L HTC, multi-scale) |
51 | – | – | – | – | – | – | – |
|
CBNet: A Composite Backbone Network Architecture for Object Detection
|
PaddlePaddle/PaddleDetection · shinya7y/UniverseNet · VDIGPKU/CBNetV2
· +1 |
2021 |