HRNet with Custom Dataset

Training High-Resolution Networks for human pose estimation on a custom COCO-style keypoint dataset

A fork of the official HRNet (High-Resolution Network) implementation, with the instructions and code changes needed to train it for human pose estimation on your own dataset.

The repository covers:

  • organizing a custom dataset in COCO format (images plus train.json / val.json keypoint annotations);
  • adapting the configuration for a different number of keypoints (12 instead of COCO’s 17);
  • training and running inference on images and videos with the resulting checkpoint.

Results on the custom dataset (374 images: 296 train, 78 validation). The best configuration, HRNet-W48 at 384×288 with ImageNet pretraining, reached 0.568 AP, 0.950 AP50 and 0.665 AR. These are below the numbers reported on COCO, which is expected given the dataset size, but good enough for practical pose estimation at inference time.

Backbone Pretrain Input size AP AP50 AP75 AR
HRNet-W32 Y 384×288 0.396 0.826 0.338 0.509
HRNet-W48 N 384×288 0.526 0.831 0.537 0.581
HRNet-W48 Y 384×288 0.568 0.950 0.579 0.665

Links: Code · Original HRNet