search

uci-uav-forge / uavf_2024

MIT License
2 stars 0 forks source link

Remove tiling and use last year's model weights, recall/precision increases by >20% on IRL data #181

Closed EricPedley closed 2 months ago

EricPedley commented 2 months ago

Summary

image Previously the precision and recall were like 50%. The shape classification accuracy went down from 80% to 60% but this still seems fine.

TBH I have no idea why this works. I was running inference to generate visualizations for the FRR and our model was making very inconsistent-looking video sequences, and I remembered when @Dat-Bois tried running last year's model with the yolov8 CLI on one of our IRL videos and it looked really good, so I just tried that and somehow we get god-tier bounding box performance. The reason I didn't notice this sooner is because when I tried this previously I kept tiling on, but this time it struck me that I should measure it quantitatively with tiling off.

Just look at this and nitpick me on code cleanliness and stuff cause I'm too tired to review myself rn.

Dat-Bois commented 2 months ago

Closing this since the code seems fine, minor changes after all.