search

XiaRho / CMDA

[ICCV23] Official Implementation of CMDA: Cross-Modality Domain Adaptation for Nighttime Semantic Segmentation
27 stars 4 forks source link

Would you consider release the pseudo labels predicted by your model for DSEC_Night training split? #5

Open Candy-Crusher opened 2 months ago

Candy-Crusher commented 2 months ago

As titled. Since "All experiments are conducted on a Tesla A100 GPU", and due to the limitation of the GPU I can get accessed to, it's hard to reproduce this work.

XiaRho commented 2 months ago

Hi, @Candy-Crusher. We understand the constraints you're facing due to GPU limitations. To generate the pseudo labels, you can download the trained model CMDA for Cityscapes→DSEC.

Once you have the model, you can follow the inference process similar to that of Cityscapes to DarkZurich python my_run_experiments.py --root_path /path_to_CMDA/CMDA/ --base_config configs/fusion/cs2dz_image+raw-isr_b5.py --name cmda_cs2dz. Modify a little code and apply this model to the DSEC-Night training set to generate pseudo labels. Rest assured, this process should not require an excessive amount of GPU memory.

Candy-Crusher commented 1 month ago

Hi! I was using python my_test.py --work_dir ./CMDA/pretrained/CMDA_DSEC_test_60.05/230301_1700_cs2dsec_image+events_together_A01B0005C1_b5_e978d but got all key in source state_dict unexpected. Could you plz give me some hints?

image

Candy-Crusher commented 1 month ago

My reproduce result is like this: image

Which seems to be align with CMDA(I) but the final miou is different?

1726044751607
XiaRho commented 1 month ago

As we said in the README, the logs provide the mIoU for 19 classes. For Cityscapes→DSEC, it is necessary to convert the mIoU to the 18 valid classes (without the truck class), i.e., the final mIoU 54.72 should be converted to 54.72*19/18=57.76.

Candy-Crusher commented 1 month ago

I see, thx for the quick respond! So the key in source state_dict unexpected can be safely ignored right?

XiaRho commented 1 month ago

Yes, there is no need for EMA parameter when testing.

Candy-Crusher commented 1 month ago

Got it. The code is clear and well-organized, thx for the great work and quick respond!

Candy-Crusher commented 1 month ago

image May I ask how did you choose the labels_range and why you substract 13 from each of them?

XiaRho commented 1 month ago

We took the first part of each sequence as labels in a certain proportion. "substract 13" seems to be related to the selection interval of the labels but is not important.