OV release DeformableConvolution support for DEKR

README.md

OV release DeformableConvolution support for DEKR

Pytorch layer documentation

OpenVINO layer documentation

To get OpenVINO IR:

To get OV model and compare the results of OV and Torch models run:

python3 tools/convert.py --cfg path/to/file.yaml TEST.MODEL_FILE path/to/file.yaml

• You can find all the changes needed to support the DeformableConvolution layer here.
• Tested with OpenVINO 2021.4 release

Expected difference:

1. For COCO:

python3 tools/convert.py \
    --cfg experiments/coco/inference_demo_coco.yaml \
    TEST.MODEL_FILE model/pose_coco/pose_dekr_hrnetw32_coco.pth

the result is:

Heatmap range: [-0.015603437, 0.036487162]
Offset range: [-44.577393, 42.865826]

Heatmap loss: 6.3575135e-07
Offset loss: 0.0032014847

2. For CrowdPose:

python3 tools/convert.py \
    --cfg experiments/crowdpose/inference_demo_crowdpose.yaml \
    TEST.MODEL_FILE model/pose_crowdpose/pose_dekr_hrnetw32_crowdpose.pth

the result is:

Heatmap range: [-0.014327541, 0.02162216]
Offset range: [-76.367546, 55.62034]

Heatmap loss: 1.206994e-06
Offset loss: 0.0017137527