Model Zoo

Common Settings

All COCO models were trained on coco_2017_train and evaluated on coco_2017_val.
All models were trained using distributed training.
Most models were trained with 50 epochs settings (~51 COCO epochs) with multi-step LR scheduler which is the common setting in DETR-like methods.

COCO Object Detection Baselines

Here we provides our pretrained baselines with detrex. And more pretrained weights will be released in the future version. We also provide our converted pretrained weights for the users which will be marked as (converted).

DETR

Name	Backbone	Pretrained	Epochs	box AP	Download
DETR-R50 (converted)	R-50	IN1k	500	42.0	model
DETR-R50-DC5 (converted)	R-50	IN1k	500	43.4	model
DETR-R101 (converted)	R-101	IN1k	500	43.5	model
DETR-R101-DC5 (converted)	R-101	IN1k	500	44.9	model

Deformable-DETR

Name	Backbone	Pretrained	Epochs	box AP	Download
Deformable-DETR + Box Refinement	R50	IN1k	50	47.0	model
Deformable-DETR + Box Refinement + Two Stage	R50	IN1k	50	48.2	model

Anchor-DETR

Name	Backbone	Pretrain	Epochs	box AP	download
Anchor-DETR-R50	R-50	IN1k	50	41.9	model
Anchor-DETR-R50 (converted)	R-50	IN1k	50	42.2	model
Anchor-DETR-R50-DC5 (converted)	R-50	IN1k	50	44.2	model
Anchor-DETR-R101 (converted)	R-101	IN1k	50	43.5	model
Anchor-DETR-R101-DC5 (converted)	R-101	IN1k	50	45.1	model

Conditional-DETR

Name	Backbone	Pretrain	Epochs	box AP	download
Conditional-DETR-R50	R-50	IN1k	50	41.6	model
Conditional-DETR-R50-DC5 (converted)	R-50-DC5	IN1k	50	43.8	model
Conditional-DETR-R101 (converted)	R-101	IN1k	50	43.0	model
Conditional-DETR-R101-DC5 (converted)	R-101-DC5	IN1k	50	45.1	model

DAB-DETR

Name	Backbone	Pretrained	Epochs	box AP	Download
DAB-DETR-R50	R50	IN1k	50	43.3	model
DAB-DETR-R50-3patterns (converted)	R-50	IN1k	50	42.8	model
DAB-DETR-R50-DC5 (converted)	R-50	IN1k	50	44.6	model
DAB-DETR-R50-DC5-3patterns (converted)	R-50	IN1k	50	45.7	model
DAB-DETR-R101	R101	IN1k	50	44.0	model
DAB-DETR-R101-DC5 (converted)	R-101	IN1k	50	45.7	model
DAB-DETR-Swin-T	Swin-Tiny-224	IN1k	50	45.2	model
DAB-Deformable-DETR-R50	R50	IN1k	50	49.0	model
DAB-Deformable-DETR-R50-Two-Stage	R50	IN1k	50	49.7	model

DN-DETR

Name	Backbone	Pretrained	Epochs	box AP	Download
DN-DETR-R50	R50	IN1k	50	44.7	model
DN-DETR-R50-DC5 (converted)	R50	IN1k	50	46.3	model

DINO

Pretrained DINO with ResNet Backbone

Name	Backbone	Pretrained	Epochs	Denoising Queries	box AP	Download
DINO-R50-4scale	R50	IN1k	12	100	49.2	model
DINO-R50-4scale (hacked trainer)	R-50	IN1k	12	100	49.4	model
DINO-R50-4scale with EMA	R-50	IN1k	12	100	49.4	model
DINO-R50-5scale	R50	IN1k	12	100	49.6	model
DINO-R50-4scale	R50	IN1k	12	300	49.5	model
DINO-R50-4scale	R50	IN1k	24	100	50.6	model
DINO-R101-4scale	R101	IN1k	12	100	50.0	model

Pretrained DINO with Swin-Transformer Backbone

Name	Backbone	Pretrained	Epochs	Denoising Queries	box AP	Download
DINO-Swin-T-224-4scale	Swin-Tiny-224	IN1k	12	100	51.3	model
DINO-Swin-T-224-4scale	Swin-Tiny-224	IN22k to IN1k	12	100	52.5	model
DINO-Swin-S-224-4scale	Swin-Small-224	IN1k	12	100	53.0	model
DINO-Swin-B-384-4scale	Swin-Base-384	IN22k to IN1k	12	100	55.8	model
DINO-Swin-L-224-4scale	Swin-Large-224	IN22k to IN1k	12	100	56.9	model
DINO-Swin-L-384-4scale	Swin-Large-384	IN22k to IN1k	12	100	56.9	model
DINO-Swin-L-384-5scale	Swin-Large-384	IN22k to IN1k	12	100	57.5	model
DINO-Swin-L-384-4scale	Swin-Large-384	IN22k to IN1k	36	100	58.1	model
DINO-Swin-L-384-5scale	Swin-Large-384	IN22k to IN1k	36	100	58.5	model

Pretrained DINO with FocalNet Backbone

Name	Backbone	Pretrained	Epochs	Denoising Queries	box AP	Download
DINO-FocalNet-Large-4scale	FocalNet-384-LRF-3Level	IN22k	12	100	57.5	model
DINO-FocalNet-Large-4scale	FocalNet-384-LRF-4Level	IN22k	12	100	58.0	model
DINO-FocalNet-Large-5scale	FocalNet-384-LRF-4Level	IN22k	12	100	58.5	model

Pretrained DINO with ViTDet Backbone

Name	Backbone	Pretrained	Epochs	Denoising Queries	box AP	Download
DINO-ViTDet-Base-4scale	ViT	IN1k, MAE	12	100	50.2	model
DINO-ViTDet-Base-4scale	ViT	IN1k, MAE	50	100	55.0	model
DINO-ViTDet-Large-4scale	ViT	IN1k, MAE	12	100	52.9	model
DINO-ViTDet-Large-4scale	ViT	IN1k, MAE	50	100	57.5	model

H-Deformable-DETR

Name	Backbone	Pretrained	Query	Epochs	box AP	Download
H-Deformable-DETR-R50 + tricks (detrex)	R50	IN1k	300	12	49.1	model
H-Deformable-DETR-R50 + tricks (converted)	R50	IN1k	300	12	48.9	model
H-Deformable-DETR-R50 + tricks (converted)	R50	IN1k	300	36	50.3	model
H-Deformable-DETR-Swin-T + tricks (converted)	Swin-Tiny	IN1k	300	12	50.6	model
H-Deformable-DETR-Swin-T + tricks (converted)	Swin-Tiny	IN1k	300	36	53.5	model
H-Deformable-DETR-Swin-L + tricks (converted)	Swin-Large	IN22k	300	12	56.2	model
H-Deformable-DETR-Swin-L + tricks (converted)	Swin-Large	IN22k	300	36	57.5	model
H-Deformable-DETR-Swin-L + tricks (converted)	Swin-Large	IN22k	900	12	56.4	model
H-Deformable-DETR-Swin-L + tricks (converted)	Swin-Large	IN22k	300	36	57.5	model

DETA

Name	Backbone	Pretrained	Epochs	box AP	Download
Improved-Deformable-DETR-R50 (converted)	R-50	IN1k	50	49.8	model
DETA-R50-5scale (bs=8, 180000 iterations)	R-50	IN1k	12	50.0	model
DETA-R50-5scale (with hacked train engine)	R-50	IN1k	12	49.9	model
DETA-R50-5scale-12ep (no frozen backbone)	R-50	IN1k	12	50.2	model
DETA-R50-5scale (converted)	R-50	IN1k	12	50.1	model
DETA-Swin-Large-finetune (converted)	Swin-Large-384	Object 365	24	62.9	model