search

mjq11302010044 / RRPN

Arbitrary-Oriented Scene Text Detection via Rotation Proposals (TMM 2018)
Other
436 stars 142 forks source link

Code performs worse on ICDAR 2015 than in paper? #11

Open idefix92 opened 6 years ago

idefix92 commented 6 years ago

Hello, I am evaluating the code on the official rrc webpage but the results are worse than in the paper. Is the code made by the authors or externally? Any reason for the decreased performance? :)

running on GPU.

Best Regards Valentin

mjq11302010044 commented 6 years ago

@idefix92 You can try testing image in a larger size. Detail information of performance can be seen in our paper on arxiv.org

shaowy commented 6 years ago

@idefix92 Please configure lib/fast_rcnn/config.py according to our paper. I found current one is not ICDAR 15. C.TRAIN.SCALES , C.TRAIN.MAX_SIZE,__C.TEST.RATIO_GROUP are of the most important if you want to use the code to train models on other datasets. If it possible, I recommend you to read lib/roi_data_layer/minibatch.py, lib/rotation/generate_anchors.py, models/rrpn/VGG16/faster_rcnn_end2end/train.prototxt. When the number of anchors (scale × angle × ratio) change to cover more cases, rpn_cls_score, rpn_bbox_pred layers should be changed.

idefix92 commented 6 years ago

Hello, thank you for quick response! What do you mean by configuring the file? What to configure exactly :)

BR Valentin

mjq11302010044 commented 6 years ago

@idefix92 Sorry, we found that our anchor combination is not the same as the provided model in training. And this bug is fixed with update the lib/fast_rcnn/config.py. You can check the newest version of RRPN :)

gittigxuy commented 6 years ago

@idefix92,have you solve the problem?could you please write a tutorial about train ICDAR 2015 dataset use this repo?