A Label Noise: Effect of Identical Patches

Here, we show that false negatives that are identical to the positive - for example, patches of the sky - do not change the sign of gradient associated with the positive. Let q be the query, u be the positive, V be the set of negatives. It is easy to see that the contribution of the negatives that are identical to the positive do not reverse the sign of the positive gradient, i.e. The proposed method outperforms many supervised methods for video object segmentation, despite relying on a simple label propagation algorithm, not being trained for object segmentation, and not training on the DAVIS dataset. We also show comparisons to pretrained feature baselines with larger networks.