osvai/af-dcd

The official project website of "Augmentation-free Dense Contrastive Distillation for Efficient Semantic Segmentation" (Af-DCD for short, accepted to NeurIPS 2023).

af-dcd's Issues

Results on Tansformer-based Structures

Thanks to the authors for their contribution. I have difficulties in reproducing the distillation experiments based on Transformer architecture to achieve the recorded performance. Can you provide more training details for reference. It would be better if .sh files are provided.

PSPNet-Res101 parameter

Hi,
Thanks for your excellent work!

I would like to receive the model parameters and trains script file of PSPNet-Res101 used in paper Table 1. (b) experiment, PSPNet-Res101 -> DeepLabV3-Res18.

model's weight trained with datasets

Hello,
Thanks for your excellent work!

I'd like to debug your OmniContrastiveFeatureLoss and figure out the details. But I can't find the teacher model weights, such as deeplabv3_resnet101_camvid_best_model.pth. I would be extremely grateful if you could provide it. Or I didn't find it.

Test data

Thanks to the authors for their contribution! However, during the reproduction process, I realised that some divisions of the test data were not provided, such as the PASCAL Voc dataset. Could the authors please provide a more comprehensive test division and the corresponding procedure?

Lfd

When Lfd is removed, the mIoU is very low.

Or can you upload the models distilled without Lfd?