Trained by Michael on the lab pc.
The change in the metrics is basically that we got slightly worse on shape color on synth data, but way better on IRL data letter color.
Metrics before:
Metrics after:
Since this one was trained on images that have real backgrounds, I also wanted to see if its performance wouldn't drop significantly on images without the background removed. IMO seems to be fine, so I think going for detection instead of instance segmentation will be a worthwhile pursuit.
Without blacking out background as pre-processing:
IRL:
Synth:
Summary
Trained by Michael on the lab pc. The change in the metrics is basically that we got slightly worse on shape color on synth data, but way better on IRL data letter color.
Metrics before:
Metrics after:
Since this one was trained on images that have real backgrounds, I also wanted to see if its performance wouldn't drop significantly on images without the background removed. IMO seems to be fine, so I think going for detection instead of instance segmentation will be a worthwhile pursuit.
Without blacking out background as pre-processing: IRL:
Synth:
Before this change, without black bg: IRL:
Synth:
