Done
The new model is finally WORKING!!
Metrics:
All are averaged over all classes.
IoU=0.5: 0.95, AP with setting threshold of 0.5 to 0.95 with pace 0.05.
IoU=0.5: Pascal VOC metric (what we are using).
Area=large, medium, small corresponding to \(Area>96^2\), \(96^2>Area>32^2\) and \(32^2>Area\). By looking at these metrics, we would see if the model is good at detecting large, medium or small objects.
maxDets are the maximum predicted objects allowed.
The metric values now are based on 4 train and 4 valid data, so it is ok to have so many zeros. Hopefully we will get more reasonable results on the whole data we have.
Todo
- The plan is uploading everything to hcc and submit jobs to get preliminary results. Finger crossed for that!