Identity Alignment by Noisy Pixel Removal
Identity alignment models assume precisely annotated images manually. Human labelling is unrealistic on large sized imagery data. Detection models introduce varying amount of noise and hamper identity alignment performance. In this work, we propose to refine images by removing the undesired pixels. This is achieved by learning to eliminate less informative pixels in identity alignment. To this end, we formulate a method of automatically detecting and removing identity class irrelevant pixels in auto-detected bounding boxes. Experiments validate the benefits of our model in improving identity alignment.
READ FULL TEXT