Closed JACK-Chen-2019 closed 11 months ago
It seems that target image has 23 classes, but the code can only predict 14 classes.
The competition stipulates that the evaluation is conducted in 14 classes. For specific details, please refer to Segmentation labels in https://www.vis.xyz/shift/get_started/
I saved the "result" from the model prediction, and it seems that the pixel labels predicted by the model are different from the actual labels, which results in a significant difference between the results submitted in ICCV2023 Workshop and the results of local testing. Is there such a problem? How to solve it?