Skip to content

Latest commit

 

History

History
26 lines (17 loc) · 1013 Bytes

20241108.md

File metadata and controls

26 lines (17 loc) · 1013 Bytes

Information

  • Time: $13:20 -$
  • Attendees: Bob Zhang (Supervisor), Huang Yanzhen, Mai Jiajun

Meeting Agenda

1. Optimized backside recognition

Deal to the design of our MLP model, we need to specify the backside of an input. The MLP model is not trained to classify people facing back to camera.

  • We defined a possible factor that can help us to know if a person is facing back to camera: $$\mathbf{backside_ratio} = \frac{\text{shoulder_x_diff}}{\text{width_diff}}$$
  • when $\mathbf{backside_ratio} > 0$, the person is facing to camera, and $<0$ means he/she is back to camera.
  • A threshold is set to be $-0.2$, a person will be classified as "Backside" if and only if $\mathbf{backside_ratio} < -0.2$.

2. Adopted new video streaming method

3. Brainstorming on Improvement

  • Do we need to improve the current models?
  • We are supposed to figure out the solution of wrong classification on a person wearing thick. (More data and analysis)

Agenda for the next meeting