CRAFT-Reimplementation
CRAFT-Reimplementation copied to clipboard
backtime92
Synth data training problem
Hi, I tried your code trainSyndata.py on data http://www.robots.ox.ac.uk/~vgg/data/scenetext/. However, after 16 epochs, the hmean on IC2013 still around 65%, which is far below your results(76.33%). Could you suggest what the possible problems are, please?
@zhenx I used the old gaussian map method to train syndata, I found the new gaussian map method pre-train model is not better than old's. So I think you can just use my pre-train model to train MLT or IC15 data. And if you need, I can share you the old gaussian map method.
@backtime92 I'm confused. So if I want to get the 82% Hmean result on ICDAR15 from scratch. I should first train the pretrained model on SynthText data with the old gaussian method (which is the same as the model you provided: link), and then finetune it on ICDAR15 using the new gaussian method?
@ddm-456 yes,and if you use the scratch from the new gaussian method, it can be ~81%. But i think you can just use the syndata pre-trained model which i had given the link.
@backtime92 Thanks for your kind reply. Yes, could you please share your old version of the heatmap code? I am wondering what the main difference between these two versions is. Could you elaborate a little bit? One more question is about your note in the readme "Note:We have solved the problem about detecting big word.". How did you deal with that problem of missing big words? What's the main action you took exactly? Thanks.
@zhenx I need to find it. there are so many versions code, I will check it and upload it.
Hi @backtime92 thank you for opening your reimplementation code of CRAFT. I want to retrain CRAFT model on Synthtext dataset but I cannot achieve the score you got. Could you share to me the gaussian method that you used for your training? Thanks.
@quangphammessi Sorry for you problem, the Syndata model I trained by old version code, the new is not suitable for training the syndata model, I think you can train the IC15 and MLT data model by using my Syndata pre-trained model. You can see my readme.md note.
hi @backtime92 . thank you for code and advice. I wonder that how many epoch did you train to make pretrained Syndata Model. I want to check training time to train my custom model. thank you.
@zhenx I need to find it. there are so many versions code, I will check it and upload it.
hi @backtime92 . did you upload it ?
