Beom

After check the pretrained-weight, I will share the link on README. The `DCNv2` is from https://github.com/CharlesShang/DCNv2, but it is no longer used (required only for `DLA_dcn.py` or `resnet_dcn.py`, hourglass does...

you should run training code in the 3 step sequentially. [please check "TextBoxes++" paper](https://arxiv.org/pdf/1801.02765.pdf)

If you look at "Model.py", it seems to be helpful. def ctc_lambda_func(args): y_pred, labels, input_length, label_length = args # the 2 is critical here since the first couple outputs of...

Are you talking about the [DeepLabV3 code](https://github.com/qjadud1994/DRS/blob/866f0e799a4e2a8b7d84016ee604f9bb0ce7f572/DeepLabV3Plus-Pytorch/eval.py#L75)? As I mentioned, we strictly followed the DeepLab-V3+ pytorch implementation from https://github.com/VainF/DeepLabV3Plus-Pytorch, and trained with our pseudo labels.

For DeepLab-V2, we did not use Resize_eval. The result of DeepLab-V3 is our additional experiment after publishing, so we skip the test-mIoU and report only the val-mIoU following the [DeepLabV3...

Initialization does not seem to matter. Because I had trained the CRNN from scratch.

Of course, if you train only with synthetic images, it is natural that it does not work properly in real images. So, I recommend pre-training with synthetic images and fine-tune...

https://github.com/qjadud1994/CRNN-Keras/issues/22#issuecomment-465437365

Of course it is possible. Just add a downsample(maxpooling) to get the output of 32-d or just use the output of 64-d.

Thank you for pointing out it! It is a bug :( We updated the code to be `if best_loss >= val_loss` and `if best_dist >= val_dist`.