Mmdetection: The new dataset Objects365 issue

[runzeer]

Hi，have you tried the recent datasets Objects365? I found it similar with the COCO dataset. So I just modified the num_classes from 81 to 366, changed the image dir as well as the annotation file and the others is the same. I use the cascade rcnn config. But when I trained the model after 1 epoch ,the validation result is zero. But the loss is gradually down. It looked abnormal. Anything else should I do?
Could you give me some advice?

[thangvubk]

To my knowledge when the number of classes is high, your score can be very low. Just try to train it more to see if the result increases or not.

[runzeer]

Thanks for your reply! But the datasets have more images than COCO. One epoch should cost more time. So is there anything wrong with my settings?

[thangvubk]

In that case, i'm not sure. Just wait if someone can answer you.

[zimenglan-sysu-512]

hi @runzeer
how do u download the Objects365?

[gxd1994]

Hi，have you tried the recent datasets Objects365? I found it similar with the COCO dataset. So I just modified the num_classes from 81 to 366, changed the image dir as well as the annotation file and the others is the same. I use the cascade rcnn config. But when I trained the model after 1 epoch ,the validation result is zero. But the loss is gradually down. It looked abnormal. Anything else should I do?
Could you give me some advice?

i have same problem

[nemonameless]

@ @runzeer @gxd1994 Have you train normally and converge? What is your epochs, bs, lr setting? Thanks

[rumusan]

I think the reason is that the order of 'categories' list in objects365_train.json an objects365_val.json is inconsistent.

And the CocoDataset in mmdetection, the label map depends on the order of 'categories' list.

self.cat_ids = self.coco.getCatIds(catNms=self.CLASSES) 
self.cat2label = {cat_id: i for i, cat_id in enumerate(self.cat_ids)}

Thus, the label in training and testing is inconsistent...

You can just modify the objects365_val.json or objects365_val.json, or the CocoDataset :

self.cat_ids = self.coco.getCatIds(catNms=self.CLASSES) 
self.cat2label = {cat_id: i for i, cat_id in enumerate(sorted(self.cat_ids))}

[sunpeng1996]

i have same problem