lightweight-human-pose-estimation.pytorch
lightweight-human-pose-estimation.pytorch copied to clipboard
Custom Dataset Annotation format
Hi,
I have a few questions after reading the tips and issues pointed out here. I am currently using Labelme to annotate the keypoints and would convert them to the COCO format later on.
-
What if my custom dataset only has keypoints annotated?
-
If i have the 18 keypoints like shown in the image below, where can i change the keypoint definitions in this repo to match my usage?
-
And for Validation of the model, are other metrics available like PCK, PCP (more info)? Would i be able to use your validation metrics with just keypoints in my GTs?
Thanks!
Originally posted by @JJLimmm in https://github.com/Daniil-Osokin/lightweight-human-pose-estimation.pytorch/issues/231#issuecomment-1207600173
Hi! It is ok, you can convert this format to COCO annotation format. Just throw away the neck and reorder keypoints as in COCO dataset, you can check #134 or the official COCO site for the COCO keypoints order. You can check additional metric in the neighborhood repository. It has the code for a single-person pose estimation.
Thanks @Daniil-Osokin . For my first pointer i think you missed that out by accident, but is it okay if i only have keypoints for my annotations? (no bboxes or segmentations)
To use the code without modifications, you need to convert the data to COCO format. You can take just bounding box around person keypoints. Segmentation, masks are not necessary, check the #231.
I see, so i can do without segmentation masks, but i would still require bounding boxes with my keypoints, so as to use it with COCO's evaluation metrics.
Great, that it is clear now!