DBNet.pytorch
DBNet.pytorch copied to clipboard
A pytorch re-implementation of Real-time Scene Text Detection with Differentiable Binarization
https://github.com/WenmuZhou/DBNet.pytorch/blob/678b2ae55e018c6c16d5ac182558517a154a91ed/base/base_dataset.py#L53 作者你好,请问__getitem__函数里,我理解是,如果出错,就重新选一张图,直到没出错为止。为什么要这么做呢?有哪些潜在的出错? 训练的时候,偶然且不规律会在这里报错,请问是为什么呢 https://github.com/WenmuZhou/DBNet.pytorch/blob/master/data_loader/modules/make_border_map.py#L51 IndexError: list index out of range
训练时在yaml中设置图片预处理的方式如下 pre_processes: - type: EastRandomCropData args: size: max_tries: 5 结果进程总是被强制kill(sigkill:9)
大佬您好,我环境是双卡,15G显存,使用6万张图片,batch_size设置成8,numworks设置为0,能正常训练,设置其他的都要报错。请问这是什么原因导致的呢?  
In my opinion, your code is better than the official code. (in readability) thanks!
 Have you done it yet??? @WenmuZhou
cosin = (square_distance - square_distance_1 - square_distance_2) / (2 * np.sqrt(square_distance_1 * square_distance_2)) square_sin = 1 - np.square(cosin) square_sin = np.nan_to_num(square_sin) result = np.sqrt(square_distance_1 * square_distance_2 * square_sin / square_distance)...
RT...原文中是两个的cb loss,我尝试把binary换成cb发现在验证集精度一致是0,很奇怪,作者是实验后换的loss还是出于什么考虑呢
感谢作者开源这个项目,我现在拿您的代码使用ic15,1c19,1c17和天猫数据训练了网络,100个epoch。准确召回如下: **recall: 0.586170, precision: 0.871295, hmean: 0.700843, train_loss: 0.759117。** 但是在检测的时候发现,文字框对于文字不能很好的包络,  请问这种情况是是我的训练epoch不够,还是有可供调节的参数呢?期待您的回复。
亲爱的作者,我在命令行下运行predict.sh,增加polygon参数,结果报错了呢,但是不赋予polygon参数就能够正常运行,我的命令如下:CUDA_VISIBLE_DEVICES=0 python3 tools/predict.py --model_path ./output_resize/DBNet_resnet18_FPN_DBHead/checkpoint/model_best.pth --input_folder ./datasets/train/img --output_folder ./test_output --thre 0.4 --polygon False --show
Thank you for your great job and share to us! Is there any way to load pretrained models from MhLiao/DB(https://github.com/MhLiao/DB)? I want to FT MhLiao/DB's models with my data.