README.md

Deformable Convolutional Networks

Introduction

@inproceedings{dai2017deformable,
  title={Deformable Convolutional Networks},
  author={Dai, Jifeng and Qi, Haozhi and Xiong, Yuwen and Li, Yi and Zhang, Guodong and Hu, Han and Wei, Yichen},
  booktitle={Proceedings of the IEEE international conference on computer vision},
  year={2017}
}

@article{zhu2018deformable,
  title={Deformable ConvNets v2: More Deformable, Better Results},
  author={Zhu, Xizhou and Hu, Han and Lin, Stephen and Dai, Jifeng},
  journal={arXiv preprint arXiv:1811.11168},
  year={2018}
}

Results and Models

Backbone	Model	Style	Conv	Pool	Lr schd	Mem (GB)	Inf time (fps)	box AP	mask AP	Config	Download
R-50-FPN	Faster	pytorch	dconv(c3-c5)	-	1x	4.0	17.8	41.3		config	model | log
R-50-FPN	Faster	pytorch	mdconv(c3-c5)	-	1x	4.1	17.6	41.4		config	model | log
*R-50-FPN (dg=4)	Faster	pytorch	mdconv(c3-c5)	-	1x	4.2	17.4	41.5		config	model | log
R-50-FPN	Faster	pytorch	-	dpool	1x	5.0	17.2	38.9		config	model | log
R-50-FPN	Faster	pytorch	-	mdpool	1x	5.8	16.6	38.7		config	model | log
R-101-FPN	Faster	pytorch	dconv(c3-c5)	-	1x	6.0	12.5	42.7		config	model | log
X-101-32x4d-FPN	Faster	pytorch	dconv(c3-c5)	-	1x	7.3	10.0	44.5		config	model | log
R-50-FPN	Mask	pytorch	dconv(c3-c5)	-	1x	4.5	15.4	41.8	37.4	config	model | log
R-50-FPN	Mask	pytorch	mdconv(c3-c5)	-	1x	4.5	15.1	41.5	37.1	config	model | log
R-101-FPN	Mask	pytorch	dconv(c3-c5)	-	1x	6.5	11.7	43.5	38.9	config	model | log
R-50-FPN	Cascade	pytorch	dconv(c3-c5)	-	1x	4.5	14.6	43.8		config	model | log
R-101-FPN	Cascade	pytorch	dconv(c3-c5)	-	1x	6.4	11.0	45.0		config	model | log
R-50-FPN	Cascade Mask	pytorch	dconv(c3-c5)	-	1x	6.0	10.0	44.4	38.6	config	model | log
R-101-FPN	Cascade Mask	pytorch	dconv(c3-c5)	-	1x	8.0	8.6	45.8	39.7	config	model | log
X-101-32x4d-FPN	Cascade Mask	pytorch	dconv(c3-c5)	-	1x	9.2		47.3	41.1	config	model | log

Notes:

• dconv and mdconv denote (modulated) deformable convolution, c3-c5 means adding dconv in resnet stage 3 to 5. dpool and mdpool denote (modulated) deformable roi pooling.
• The dcn ops are modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch, which should be more memory efficient and slightly faster.
• (*) For R-50-FPN (dg=4), dg is short for deformable_group. This model is trained and tested on Amazon EC2 p3dn.24xlarge instance.
• Memory, Train/Inf time is outdated.