search

yanxp / MetaR-CNN

Meta R-CNN : Towards General Solver for Instance-level Low-shot Learning
https://yanxp.github.io/metarcnn.html
179 stars 23 forks source link

I found un-fair data usage #33

Open mandal4 opened 4 years ago

mandal4 commented 4 years ago

Hi there. I well read uploaded source code. I found you set k-shot masked sample with k-'image' not 'instance'. It might be okay that there is binary mask for input of PRN. But the same k-'image' is fed into fasterRCNN in 2nd phase. So for FasterRCNN, it could be k+1 shot or k+2 shot and so on because you set k-shot with image-wise. Unfortunately i think it is un-fair setting for few-shot learning. Could you explain about it? I hope i misunderstood your nice work.

yanxp commented 4 years ago

Hello, we have filtered the images with k-shot instance. Please see the function https://github.com/yanxp/MetaR-CNN/blob/0e54c48505a0fd472eec3885d1ea2a80852cf681/lib/roi_data_layer/roidb.py#L61

mandal4 commented 4 years ago

Thanks for reply. I confused that. But still i couldn't understand it is same for input of PRN

Ze-Yang commented 4 years ago

I have the same concern with you and I have create an issue on the top. Maybe you can check that. I think I express the same idea with you.

Ze-Yang commented 4 years ago

Hi there. I well read uploaded source code. I found you set k-shot masked sample with k-'image' not 'instance'. It might be okay that there is binary mask for input of PRN. But the same k-'image' is fed into fasterRCNN in 2nd phase. So for FasterRCNN, it could be k+1 shot or k+2 shot and so on because you set k-shot with image-wise. Unfortunately i think it is un-fair setting for few-shot learning. Could you explain about it? I hope i misunderstood your nice work.

I also find weird modeling. After the reweighted feature, it also produce num_classes+1 prediction. So what's the contribution of feature reweighting?