mhruz / POTR

Pose Transformer
3 stars 1 forks source link

How to use this code to NYU/icvl hands dataset? #5

Open pijiangdada opened 1 year ago

pijiangdada commented 1 year ago

i read the paper(MuTr: Multi-Stage Transformer for Hand Pose Estimation from Full-Scene Depth Image), and got this code, but i found it difficult to use the code to NYU/icvl dataset, may i ask if there has README.md about how to apply the model in hand pose dataset?

mhruz commented 1 year ago

Hi, it is true, that we expect a H5 file with depth data and normalized labels (-1. 1). https://github.com/mhruz/POTR/blob/eafbe010268aa9e0bcef10ae2cd2314fe199d9fc/dataset/hpoes_dataset.py#L158

This is not described here, but we provide the information in the paper. The codes that load the NYU dataset are not in a "presentable" state. However, I will check with my colleagues who worked on it whether they can share the codes for that.

Furthermore, for datasets such as NYU it might be enough to use the deformable POTR model without the multi-stage approach which is more suitable when the hand region is not known.

zdkrnoul commented 1 year ago

The codes that load the NYU and others datasets are in directory ./utils/