[WIP] Support Mask R-CNN w/ keypoints

[yuyu2172]

Merge after https://github.com/chainer/chainercv/pull/793 and https://github.com/chainer/chainercv/pull/781

[yuyu2172]

With this weight ~https://drive.google.com/open?id=1UJ5LMwiyK-k1g-bv-4dmiSKej7LA_2NV~, demo works https://drive.google.com/open?id=1UJ5LMwiyK-k1g-bv-4dmiSKej7LA_2NV

$ python3 demo.py --gpu 0 --model mask_rcnn_fpn_resnet50 --pretrained-model chainercv_point_rcnn_fpn_resnet50_coco_converted.npz  --mode keypoint IMG_PATH.jpg

Evaluation

converted (bilinear interpolation with kernel size 2): 61mmAP 62cb5e8 converted (bilinear interpolation with kernel size 4): 64.0mmAP 9a26065 caffe2: 64.2mmAP (https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md)

[yuyu2172]

TODOs

• [x] Use balanced sampling to include negative samples for keypoint head loss
• [x] use within_bbox to decide invalid boxes
• [x] multi scale training

[yuyu2172]

Transition is working at least for inference.

mmAP (all): 0.639657
mmAP (large): 0.73500156
mmAP (medium): 0.58247757

[yuyu2172]

python3 eval_keypoint_detection.py --model keypoint_rcnn_fpn_resnet50 --pretrained-model coco --gpu 0

mmAP (all): 0.6396571 mmAP (large): 0.73500156 mmAP (medium): 0.58247715

このときと https://github.com/chainer/chainercv/pull/810#issuecomment-473249333 同じ