Open isaacperez opened 10 months ago
I set category_space=None
and that was the problem, doesn't the trained model work without creating the file for category_space
?
i download the files below and run the demo command without anything change,i get the output like the second picture,it seems normal
That means you have used category_space="demo/ycb_prototypes.pth"
. So you have overwritten the model classes. If you leave category_space=None
, the labels look random, shouldn't they be COCO or LVIS classes?
That means you have used
category_space="demo/ycb_prototypes.pth"
. So you have overwritten the model classes. If you leavecategory_space=None
, the labels look random, shouldn't they be COCO or LVIS classes?
Hi @isaacperez , the labels seem to be from LVIS if you set category space as none. LVIS has lots of long-tailed classes.
Hi, in that case the model file does not contain the correct prototypes/classes (see the image at the beginning of the issue) because, for example, it is labeling a pringles can as a control.
Do you think that is possible with LVIS prototypes?
Do you think that is possible with LVIS prototypes?
Yes. Actually this demo is a highly challenging scenario, more difficult than COCO or LVIS I would say because of occlusion, very crowded scene, retailed categories (retail objects recognition is in itself hard) and etc.
So ideally when using LVIS prototypes, it shall not produce as many boxes, I haven't tried this either but turns out it does. This is our limitation for sure. This could very likely motivate future work on preventing false positive in open-set detection, how to handle retail categories and etc.
I have downloaded everything but when I try the demo I get the same bounding box but the wrong labels:![ycb out](https://github.com/mlzxy/devit/assets/12383906/52b01c89-0b09-4864-b528-e4ae1ab1863d)