TextBoxes_plusplus_Tensorflow
TextBoxes_plusplus_Tensorflow copied to clipboard
Shun14
测试精度问题
I use the released ckpt in this git, test on the ICDAR Incidental 2015 test set. The result seems not good enough. Recall/Precision/Hmean: 0.666/0.035/0.067 Using prob score, under the best hmean, p/r/f1-score: 0.453/0.406/0.428 I assume the author didn't train on this training set, so next i'll try and show the results.
I use the released ckpt in this git, test on the ICDAR Incidental 2015 test set. The result seems not good enough. Recall/Precision/Hmean: 0.666/0.035/0.067 Using prob score, under the best hmean, p/r/f1-score: 0.453/0.406/0.428 I assume the author didn't train on this training set, so next i'll try and show the results.
The released model was trained on SynthText Dataset.You can train your own model base on it.
The author says N in loss function is the number of default boxes that match groundtruth boxes, but in this code is the batch size instead.
Follow the author's training parameters in Table, I have the following result: p/r/f1-score: 0.773/0.697/0.733
Follow the author's training parameters in Table, I have the following result: p/r/f1-score: 0.773/0.697/0.733
Could you tell me How you get this result, have you modified the loss function or other ways.The loss value does not decreas when training on icdar2015.
Follow the author's training parameters in Table, I have the following result: p/r/f1-score: 0.773/0.697/0.733
Could you tell me How you get this result, have you modified the loss function or other ways.The loss value does not decreas when training on icdar2015.
I have the some problem when i use the released model which pre-trained on Synthtext, the loss seem ridiculous and wave between 70 to 200 (sometime even higher). Have you guys solved this problem?
