Masking errors, e.g. S_16/50.png training mask

Hi, While investigating why my predictions are not improving, I noticed an error in S_16/50.png training mask. Left side mask boundary follows the light/lens reflections, which is not correct (I checked with the rest of S_16). So, if you test mask(s) has the same error(s), the prediction/test error will remain regardless of how good a model is. Any suggestions? I can manually remove the training erroneous masks, but not the erroneous test masks.

Thanks for pointing out the error.

The masks were manually labeled and are treated as the ground truth. The errors from GT is label noise.
Besides, you could ignore the identified erroneous GT during training.