Evaluation of deformable conv nets on real-world problems about deformable-convolution-pytorch

[1zb]

No. I don't have enough computational power to do this.

[lshiwjx]

Have you measure it? @mgalkov , In my experiment I found it has little effect because it can only learn the offset near the 0.

[zhanglonghao1992]

@lshiwjx Me too.. I trained faster_rcnn on VOC 07+12 dataset, using ResNet50 with dcn in stage 5. I found that the offsets it learned during training are too small, usually 1e-2~1e-3.. It seems like the offset does not play any role.
But when training the classification network, offsets become too big. Almost all values added with offsets are out of the boundary..