Open aooating opened 2 weeks ago
Hi.
This part can help ease your concerns. https://github.com/clovaai/ECLIPSE/blob/9a7cdfbc3615537ff3b4f54af69a4974aff7608f/mask2former/maskformer_model.py#L390-L393
The index of no-obj is -1, which is the same as the number of classes.
Here, labels.ne(self.sem_seg_head.num_classes)
means filtering out no-obj masks.
(labels == -1) is equal to (labels == self.sem_seg_head.num_classes)
self.class_embed = IncrementalClassifier( [1] + classes, channels=hidden_dim, bias=bias, deep_cls=deep_cls, ) The code of classifier indicates the index of no_obj is 0, if there is an error?
It is not an error.
Please check the implementation of IncrementalClassifier
.
The first class is placed on the last index.
In logit manipulation implementation, no-obj seems like the last class, while its index is 0 in ADE20K_SEM_SEG_CATEGORIES. I want to know if there is an error, or my understanding is wrong. Looking forward to the author's reply.