ACAR Robustness for Inactive Agents

[Edwardius]

How do we make our perception stack conscious of active and inactive agents?
Some proposals so far:
Optical Flow (YOLOACT)
Video-Level detection

[liqq010]

An improvement of detection on inactive agents by assigning an additional class for agents.

Inactive agent detection achieve [email protected] of 34.7% on val1, improving about 4% than previous model.
And achieve [email protected] of 49.1% of “Car” agent, improving 4.7% than previous model.
Combined with merged class, an improvement of 1.5% on end-to-end evaluation compared to the best detector of last year.

https://github.com/WATonomous/action-classification/wiki/Agent-Detection#inactive-agent-detection