py-R-FCN icon indicating copy to clipboard operation
py-R-FCN copied to clipboard

Ignore Difficult_Case set to True

Open ananddb90 opened this issue 7 years ago • 1 comments

Hello @YuwenXiong I have a doubt regarding ignoring difficult_case while training/testing on Pascal0712.

As per my understanding, for both training and testing phase, difficult annotation from pascal data is ignored which is set in pascal_voc.py # PASCAL specific config options self.config = {'cleanup' : True, 'use_salt' : True, 'use_diff' : False, 'matlab_eval' : False, 'rpn_file' : None, 'min_size' : 2}

As per my understanding, one should not ignore difficult_case atleast for evolution because if network is able to detect difficult example from test set, it will be treated as false alarm since there is no ground truth available to evaluate.

Please correct me if my understanding is wrong.

Thank you

ananddb90 avatar Aug 17 '17 14:08 ananddb90

I'm not sure whether I understand your question, and the 'use_diff' is only used during training https://github.com/YuwenXiong/py-R-FCN/blob/master/lib/datasets/pascal_voc.py#L260.

liyi14 avatar Sep 09 '17 03:09 liyi14