search

Daniil-Osokin / lightweight-human-pose-estimation-3d-demo.pytorch

Real-time 3D multi-person pose estimation demo in PyTorch. OpenVINO backend can be used for fast inference on CPU.
Apache License 2.0
660 stars 138 forks source link

Should I retrain the model because of camera extrinsic parameters? #42

Closed mtsai101 closed 3 years ago

mtsai101 commented 3 years ago

Hi, I want to use my own camera to inference the 3D positions of the joints of the body. Do I need to retrain the model to get the correct 3D positions of joints? Or I just need to modify the extrinsics.json and run demo.py I have replaced the extrinsic.json with my own value, but I can only get reasonable points in the 2D image but not in the 3D coordinate. BTW, I get intrinsic and extrinsic parameters by OpenCV: here

If something is still not clear, please let me know. Thank you!

Daniil-Osokin commented 3 years ago

Hi! Did you try to run the demo as is (with default extrinsics)? The network uses just an image for inference, extrinsics used only for the visualization. You can check training camera positions in CMU Panoptic Dataset.

mtsai101 commented 3 years ago

Thanks for the reply! I can run the demo with the default extrinsics successfully and normally, but fail to run it with my own extrinsics to get correct 3D positions. I think probably just the approach that I get my extrinsics goes wrong. Can you share with me how you get you extrinsics? I'll very appreciate it because I've tried two to three methods.

Thanks again!

Daniil-Osokin commented 3 years ago

We are using default extrinsics from the Panoptic dataset (corresponding to front facing camera).

Daniil-Osokin commented 3 years ago

Hope, it is clear now.