MCG-NJU / SparseOcc

[ECCV 2024] Fully Sparse 3D Occupancy Prediction & RayIoU Evaluation Metric
https://arxiv.org/abs/2312.17118
Apache License 2.0
235 stars 16 forks source link

question about class balanced #7

Closed Joeless closed 4 months ago

Joeless commented 5 months ago

Hi, thanks for your code! image I noticed in the sparsev2 paper that category balancing is important, but I wonder if it is effective for dense occ method? If the current GT does not have some classes, the weight will be +inf. I tried to use it in dense OCC prediction and found that is harmful for miou.

Joeless commented 5 months ago

Another question, I don't see that you used category balance in the semantic part of your code. Has this strategy been abandoned?

afterthat97 commented 5 months ago

the class balanced CE is here:

https://github.com/MCG-NJU/SparseOcc/blob/b35e1d81d4bdbd7cc787f890f1c21885a754ac4f/models/loss_utils.py#L63

Joeless commented 5 months ago

Is this part of the code only used for 2-cls occupancy pred?

afterthat97 commented 5 months ago

yes