kohillyang / mx-detection

A mxnet object detection library contains implementations of RFCN, FCOS, RetinaNet, OpenPose, etc..

Pre-trained Models

You can download pre-trained models from https://drive.google.com/drive/folders/1LQnVHb5Xo6fKknUiOa1fXmGI_MCucGTC?usp=sharing

ModelName	Dataset	Backbone	mAP	with DCN	with Sync BN	Target Size	Max Size	IM_PER_IMAGE	Number of GPUs	Epochs
FCOS No Tricks	COCO2017	ResNet50	0.367	False	False	800	1333	4	4	6
FCOS No Tricks	COCO2017	Mobilenetv1-1.0	0.222	False	False	500	833	4	4	6
FCOS	COCO2017	ResNet50	-	True	False	800	1000	2	3	14
HRNet-cls	-	-	See HRNet	-	-	-	-	-	-	-
RetinaNet	COCO2017	ResNet50	0.325	False	False	500	833	2	3	6
OpenPose	COCO2017	Dilated-ResNet50	0.564	False	False	368	368	4	3	40
OpenPose	COCO2017	VGG16	0.561	False	False	368	368	4	3	40
RFCN	VOC12+07	Dilated-ResNet101	0.825	Only 3	False	800	1280	1	3	6
RFCN	VOC12+07	Dilated-ResNet50	0.804	Only 3	False	800	1280	1	3	6
FPN(MS)	COCO2017	SEResNext50_32x4d	0.376	True	False	800	1280	1	4	5
FPN(MS)	COCO2017	Dilated-ResNet101	0.412	True	False	800	1280	1	4	5

Notes:
FCOS No Tricks means the setting is same as original paper, i.e., centerness is on cls branch, GN is added, use P5 instead of C5, and other setting like norm_on_bbox, centerness_on_reg, center_sampling is set to False. The mAP reported by the original paper is 0.371. For more information about FCOS, please see fcos.md

For OpenPose, please go into https://github.com/kohillyang/mx-openpose for more information.

RFCN trained on VOC is reported as mAP@IoU=0.5 according to VOC Metric, and it is slightly different from mAP @IoU=0.5 of COCO.

MS means the model is using multi-scaling when training.

FPN(MS) and RFCN are bought from https://github.com/msracver/Deformable-ConvNets and rewritten by new Gluon API, their performance should be same with the model from https://github.com/msracver/Deformable-ConvNets.

Acknowledgements

Greatly thanks to https://github.com/wkcn/MobulaOP by @wkcn.

If you have any question or suggestion, please feel free to send me a mail or create an issue.

Todo List:

• FCOS+Tricks(center sampling, centerness on reg head, gIoU, gFocalLoss, etc.).
• ATSS/PAA based on FCOS.
• Train OpenPose with HRNet
• PolarMask