Hier-R-CNN

Official implementation of Hier R-CNN: Instance-level Human Parts Detection and A New Benchmark (under review)

In this repository, we release the COCO Human Parts dataset and Hier R-CNN code in Pytorch.

• Hier R-CNN architecture:

• Hier R-CNN output:

Installation

• 8 x TITAN Xp GPU
• pytorch1.1
• python3.6.8

Install Hier R-CNN following INSTALL.md.

Dataset

You can download the annotations of COCO Human Parts here. And following Data to train or evaluate Hier R-CNN models.

Results and Models

Backbone	LR	MS training	DCN	Det AP	Sub AP	DOWNLOAD
R-50-FPN	1x	No	No	36.8	20.0	GoogleDrive
R-50-FPN	2x	Yes	No	39.3	23.2
R-50-FPN	1x	No	Yes	38.6	21.9
R-50-FPN	2x	Yes	Yes	40.6	23.0	GoogleDrive
R-101-FPN	1x	No	No	37.2	20.6
X-101-FPN	1x	No	No	38.8	22.2
X-101-FPN	2x	Yes	No	40.5	24.1
X-101-FPN	1x	No	Yes	40.3	22.9
X-101-FPN	2x	Yes	Yes	42.0	24.2	GoogleDrive

ImageNet pretrained weights

• R-50
• X-101-32x8d

Training

To train a model with 8 GPUs run:

python -m torch.distributed.launch --nproc_per_node=8 tools/train_net.py --cfg cfgs/mscoco_humanparts/e2e_hier_rcnn_R-50-FPN_1x.yaml

Evaluation

multi-gpu evaluation,

python tools/test_net.py --cfg ckpts/mscoco_humanparts/e2e_hier_rcnn_R-50-FPN_1x/e2e_hier_rcnn_R-50-FPN_1x.yaml --gpu_id 0,1,2,3,4,5,6,7

single-gpu evaluation,

python tools/test_net.py --cfg ckpts/mscoco_humanparts/e2e_hier_rcnn_R-50-FPN_1x/e2e_hier_rcnn_R-50-FPN_1x.yaml --gpu_id 0

License

Hier-R-CNN is released under the MIT license.