search

VlSomers / keypoint_promptable_reidentification

[ECCV24] Keypoint Promptable Re-Identification: SOTA ReID method robust to occlusions and multi-person ambiguity
Other
87 stars 5 forks source link

How I can use my own dataset for training？ #4

Open BestTao opened 2 months ago

BestTao commented 2 months ago

Hello, I would like to ask how I can use my own dataset for training

VlSomers commented 2 months ago

Hi @BestTao , there is not tutorial yet to do so, what I would advice is to run the pipeline with a classic dataset (e.g. occluded-duke), understand the data format and how data is loaded/processed, and try to mimic it with your own dataset. The keypoints can be generated with any pose estimator of your choice or manually (I used OpenPifPaf), and the human parsing labels can be generated with a combination of PifPaf and SAM for instance. There is a script on BPBreID GitHub to generate the human parsing labels: https://github.com/VlSomers/bpbreid/blob/main/torchreid/scripts/get_labels.py. If you have good software dev skills, feel free to open a pull request with your script to generate all necessary annotations for an external dataset! Let me know if you need further details