search

CityU-AIM-Group / SOMA

[ICCV' 23 ORAL] Novel Scenes & Classes: Towards Adaptive Open-set Object Detection
MIT License
35 stars 4 forks source link

Train error after corrections in DAOD.py #6

Closed Jandsome closed 7 months ago

Jandsome commented 7 months ago

Hi, Excellent works!
Something was wrong when I was training this week. I guess it's because you corrected the DAOD.py three weeks ago . I found you corrected in line 155 two times, but the second time you changed the first correction back. I don't konw why.
And in line 45 you changed
'val_data_list': root / 'Cityscapes/AOOD_Main/val_target.txt'
to 'val_data_list': root / 'Cityscapes/AOOD_Main/train_target.txt' I wonder whether it's appropriate if train_data_list is same to val_data_list. And The error comes to me is FileNotFoundError: [Errno 2] No such file or directory: '/data/jyq/SOMA/datasets/dataset/Cityscapes/leftImg8bit_foggy/val/jena/jena_000061_000019_leftImg8bit_foggy_beta_0.02.png' FileNotFoundError: [Errno 2] No such file or directory: '/data/jyq/SOMA/datasets/dataset/Cityscapes/leftImg8bit_foggy/val/hamburg/hamburg_000000_029144_leftImg8bit_foggy_beta_0.02.png'

However,it didn't happen before you made thess corrections, it will be so kind of you to help me solve this issue. Thank you!

wymanCV commented 7 months ago

Hi, thank you for your attention! I just corrected the image folder of the dataset, which I think should work now.

https://github.com/CityU-AIM-Group/SOMA/blob/08182aec101abac275cbad857e1c9fca894bbe13/datasets/DAOD.py#L39

Using the same unlabeled target-domain dataset for training and test is reasonable for some existing DAOD benchmarks, e.g., P2C. See here from the first work: Strong-Weak Distribution Alignment for Adaptive Object Detection

image

The reason we didn't use the val set for Cityscapes is that it only contains 500 images with very limited open-set objects, which are insufficient for some open-set detection metrics, e.g., AOSE., and the number of AOSE will be very very small and unstable.

Jandsome commented 7 months ago

thanks a lot ,I think it will be work after you correct in line 39