search

yanxp / MetaR-CNN

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

Report a bug in meta training phase 2 #16

Open Ze-Yang opened 4 years ago

Ze-Yang commented 4 years ago

There is a bug in the function filter_class_roidb(roidb, shot, imdb) in line 59 of file MetaRCNN/lib/roi_data_layer/roidb.py, resulting in actual novel shots exceeding designated shots. For instance, supposing we have collect enough bounding boxes of novel class 'sofa'. At this time, if an image contain 2 bounding boxes (one for whatever base class while the other is 'sofa'), we hope that this image will be excluded, otherwise extra 'sofa' will be used in training. However, since this image contains a base class bounding box, variable flag will be turned into True, making it be included in the training data. Thus, I think this is a bug need to be fixed. Thanks.

yanxp commented 4 years ago

We have updated roidb's key and value by update_keyvalue

Ze-Yang commented 4 years ago

Actually, r[k] will never be a list. Thus, I consider nothing will be updated through https://github.com/yanxp/MetaR-CNN/blob/353b4ed25e4b2ba23442f999b3990d5b90b30345/lib/roi_data_layer/roidb.py#L54.