search

suyukun666 / UFO

Official PyTorch implementation of the “A Unified Transformer Framework for Co-Segmentation, Co-Saliency Detection and Video Salient Object Detection”. (TMM2023)
MIT License
291 stars 49 forks source link

How to deal with the noisy ground truth of COCO-SEG? #4

Open sourachakra opened 2 years ago

sourachakra commented 2 years ago

I observed that several ground truth segmentation masks provided in the COCO-SEG for certain categories are noisy i.e. the masks are several other objects are also present along with the mask of the desired object. Some examples are cup, tennis racket (humans are segmented along with the racket in the ground truth mask). How should we deal with this? Do you pre-preprocess this dataset somehow? Thanks.

suyukun666 commented 2 years ago

Thanks for paying attention to this issue. Actually, we did not take these noisy labels into consideration. And this is the main difference between co-segmentation (segment out all the co-objects) and co-saliency detection(segment out only the most salient co-objects), and thus, some other methods will use DUT dataset as we mentioned in the paper for further training to reduce the noise.
If you are interested, you can try to use additional salient datasets to improve the performance.