search

akshitac8 / OW-DETR

[CVPR 2022] Official Pytorch code for OW-DETR: Open-world Detection Transformer
231 stars 40 forks source link

OWOD data split doesn't work. #27

Open orrzohar opened 1 year ago

orrzohar commented 1 year ago

Hi, So I have been trying to implement the old TOWOD data split. I found that the images specified there are the VOC2007 classes. However, there is a bug where the classes defined at the top of "datasets/torchvision_datasets/open_world.py" are not the VOC2007 classes:

VOC_CLASS_NAMES = [ "aeroplane","bicycle","bird","boat","bus","car", "cat","cow","dog","horse","motorbike","sheep","train", "elephant","bear","zebra","giraffe","truck","person" ]

and missing: bottle, chair, dining table, potted plant, sofa, tv/monitor.

Did you train on these classes when you report the results of task 1 table 1? did you use the VOC2007 classes? if so, could you upload the scripts you used?

Thank you in advance

akshitac8 commented 1 year ago

Hello @orrzohar We are not training on bottle, chair, dining table, potted plant, sofa, tv/monitor. for the task 1 setting. These classes come in the future task.

orrzohar commented 1 year ago

Even for the results in table 1 of your MS? Because you stated that you train on PASCAL VOC classes.

Best Orr