Information

Time: 14:15-14:40
Attendees: Bob Zhang (Supervisor), Mai Jiajun, Huang Yanzhen

Discussion Summary

Code Optimization

Stress Testing: New York Times Square, performance issue
Current Test Result: Mediapipe takes up too much time
Interface the pred_instances from mmpose package with the project classifier.

Similar Work on GitHub

MMPose:
- A well-encapsulated project on GitHub.
- Use top-down predictor to predict key points.
- ~130 key points for us to choose.
OpenPose:
- Advertised to be "faster".
- Harder to use, very few key points.
- Need to study via videos.

Migration

Currently, Mediapipe takes up too much time per frame running only on CPU since Mediapipe does not support CUDA.
Gradually migrate to CUDA-supportive projects, like MMPose, OpenPose, etc.
Current Issue: MMPose does not work on CUDA12.5 with python 3.8. Need a computer/workstation with lower CUDA version to test if this issue still exists.

Push Forward: WebSocket Streaming

Python + Next.js
Push video frames with base64 encoding to WebSocket.
Frontend receives base64 and recover video frame. (WIP)

Others

Be aware of the mutual-similarity issue.

Agenda of Next Meeting

Migration of opensource pose detection projects with properly arranged interfaces.