model_zoo.md

Benchmark and Model Zoo

CNN-based (w/ ImageNet-1k pretrained)

Faster R-CNN

Backbone	Lr Schd	box mAP (minival)	#params	FLOPs	config	log	model
DB-ResNet50	1x	40.8	69M	284G	config	github	github

Cascade R-CNN (1600x1400)

Backbone	Lr Schd	box mAP (minival/test-dev)	#params	FLOPs	config	model
DB-Res2Net101-DCN	20e	53.7/-	141M	429G	config	github
DB-Res2Net101-DCN	20e + 1x (swa)	54.8/55.3	141M	429G	config (test only)	github

Cascade R-CNN w/ 4conv1fc (1600x1400)

Backbone	Lr Schd	box mAP (minival/test-dev)	#params	FLOPs	config	model
DB-Res2Net101-DCN	20e	54.1/-	146M	774G	config	github
DB-Res2Net101-DCN	20e + 1x (swa)	55.3/55.6	146M	774G	config (test only)	github

Notes:

• For SWA training, please refer to SWA Object Detection

Transformer-based (w/ ImageNet-1k pretrained)

Mask R-CNN

Backbone	Lr Schd	box mAP (minival)	mask mAP (minival)	#params	FLOPs	config	log	model
DB-Swin-T	3x	50.2	44.5	76M	357G	config	github	github

Cascade Mask R-CNN w/ 4conv1fc

Backbone	Lr Schd	box mAP (minival)	mask mAP (minival)	#params	FLOPs	config	log	model
DB-Swin-T	3x	53.6	46.2	114M	836G	config	github	github

Cascade Mask R-CNN w/ 4conv1fc (1600x1400)

Backbone	Lr Schd	box mAP (minival/test-dev)	mask mAP (minival/test-dev)	#params	FLOPs	config	model
DB-Swin-S	3x	56.3/56.9	48.6/49.1	156M	1016G	config	github

Transformer-based (w/ ImageNet-22k pretrained)

HTC (1600x1400)

Backbone	Lr Schd	box mAP (minival/test-dev)	mask mAP (minival/test-dev)	#params	FLOPs	config	model
DB-Swin-B	20e	57.9/-	50.2/-	231M	1004G	config	github
DB-Swin-B	20e + 1x (swa)	58.2/58.6	50.4/51.1	231M	1004G	config (test only)	github

HTC (bbox head w/ 4conv1fc) (1600x1400)

Compared to regular HTC, our HTC uses 4conv1fc in bbox head.

Backbone	Lr Schd	box mAP (minival/test-dev)	mask mAP (minival/test-dev)	#params	FLOPs	config	model
DB-Swin-B	20e	58.4/58.7	50.7/51.1	235M	1348G	config	github
DB-Swin-L	1x	59.1/59.4	51.0/51.6	453M	2162G	config	github
DB-Swin-L (TTA)	1x	59.6/60.1	51.8/52.3	453M	-	config	github

TTA denotes test time augmentation.

Notes:

• Pre-trained models of Swin Transformer can be downloaded from Swin Transformer for ImageNet Classification.