pytorch-faster-rcnn icon indicating copy to clipboard operation
pytorch-faster-rcnn copied to clipboard

Published 20 hours ago • verified publisher icon AlphaJia

Question about the Result

Open Hiroy101 opened this issue 4 years ago • 5 comments

I tried to use VOC dataset, so i have modified the dataloader and the evaluation code with some parts of mmedetction. However, i got really pool result (mAP is about 0.167) after 17 epochs training（about 1 day）. So, i'm wondering how many epochs did your coco dataset cost, and what's your final AP result? I would be very grateful if you could give me the answer.

Hiroy101 avatar Nov 28 '20 07:11 Hiroy101

Hello @Hiroy101 , Did you get any better results. I'm able to reproduce around 19-20 mAP on PASCAL VOC validation

deepaksinghcv avatar Feb 24 '21 04:02 deepaksinghcv

@deepakksingh, finally i switched to other frameworks cause i did not get any better result since then.

Hiroy101 avatar Feb 24 '21 04:02 Hiroy101

@Hiroy101 Thanks for your quick reply. I'm able to reproduce results using mmdetection and Detectron2. BTW, were you able to train using resnet50_fpn backbone in this repo ? It throws RuntimeError: shape '[387072, -1]' is invalid for input of size 1571328 when i try to change the backbone to resnet50_fpn.

deepaksinghcv avatar Feb 24 '21 05:02 deepaksinghcv

@deepakksingh, I seem to have encountered a similar problem at that time, it should be because the size of the FPN output feature map does not match the subsequent input, maybe you can try to modify it. this is all the help i can give u, sorry. i cant remember other specific details -_-||

Hiroy101 avatar Feb 24 '21 05:02 Hiroy101

Thanks @Hiroy101, I will check it.

deepaksinghcv avatar Feb 24 '21 05:02 deepaksinghcv