Hello @MhLiao, I would like to thank you for the great work !
But I am experiencing problems with running the network. My results differ significantly from what others archive. The network fails to detect text in several areas leading to very low recall (0.3) which in turn leads to a low f-measure (0.5). The precision is quite high - 0.9.
where the first line is yours and the second is mine. I reordered the detections.
I downloaded the icdar2015 pretrained model from dropbox and it has a sha1 sum of 4c80a5ee6c7e2eebaa200ad8b75d68a048abc6b4
I use the code from the examples/text/demo_det.py file.
Here are some failure examples, the red boxes are the ground truth ones and the green boxes are what is detected.
I also run the model on icdar focused scene text dataset the and it gives better results - F1 score: 0.708376421923 precision: 0.816448152563 recall: 0.625570776256.
Hello @MhLiao, I would like to thank you for the great work !
But I am experiencing problems with running the network. My results differ significantly from what others archive. The network fails to detect text in several areas leading to very low recall (0.3) which in turn leads to a low f-measure (0.5). The precision is quite high - 0.9.
I have compared my results on demo.jpg with yours an they are quite similar - https://github.com/MhLiao/TextBoxes_plusplus/issues/19
where the first line is yours and the second is mine. I reordered the detections.
I downloaded the icdar2015 pretrained model from dropbox and it has a sha1 sum of 4c80a5ee6c7e2eebaa200ad8b75d68a048abc6b4
I use the code from the
examples/text/demo_det.pyfile.Here are some failure examples, the red boxes are the ground truth ones and the green boxes are what is detected.
I also run the model on icdar focused scene text dataset the and it gives better results - F1 score: 0.708376421923 precision: 0.816448152563 recall: 0.625570776256.