Closed hongzhidu closed 2 years ago
In our experiments, the semantic id for each object was fixed during training, i.e., the objects of the same class had the same semantic id. I am curious, in which case that the semantic label of the object would be not fixed during training.
Thanks for your reply! I just think the semantic id is unnecessary in this task. It might be more appropriate to do instance segmentation.
In fact, semantic labels are common in these tasks, such as semantic segmentation for background and foreground point clouds in the stacked scene.
Thanks again for your reply!
Hi, Thanks for your excellent work!
May I consult you about why use cross entropy as the loss of semantic segmentation? If the label of each object is not fixed during training, will this loss function be invalid?