SidneyRey

question about training with crop image as input

2

作者您好，我尝试用center crop的图像作为训练的输入。我改动了如下内容： 1）Image_base中get_item_single_frame，对目标中心做center crop。 2）all_person_detected_mask，所有sample都是true 3）max_person为1，multi_person为false 4）训练集h36m,mpiinf,mpii,lsp,crowdpose,pw3d 用pretrained model finetune loss中MPJPE和PAMPJPE基本不降(维持在200左右)，训练图片可视化中，能看到mpiinf预测结果越来越差，训练几个epoch后这个数据集的结果就完全不对了，但其他数据集看上去没什么问题。 然后，我将mpiinf数据集去掉训练，loss和可视化结果看上去都正常。 整图训练，用mpiinf，MPJPE和PAMPJPE是在下降的，但是crop图时，只有这个数据集降不下来，您知道原因不?

hello, I want to train widerface datasets, I have two Q. 1) where to download file widerface2019_train_cocostyle.json? 2) what the dataset structure for COCO? thank you !

about pretrained model

2

I got an error after run ./scripts/download_model.sh 387. requests.exceptions.MissingSchema: Invalid URL '1LS_Ocw8LjHAlBalHNtJJSS0aHJqc3TKR': No schema supplied. Perhaps you meant http://1LS_Ocw8LjHAlBalHNtJJSS0aHJqc3TKR? How to download the pretrained model? Thank you for your work.

Thank you for your amazing work! I do not know how to animate avatar, could you share the tools or code about it.

Thank you for your wonderful work! I want to test your work in my data. My data is in Markdown format, with tables and formulas in LaTeX. But, I do...

TextDetection每页开销1~4秒？

1

请问下，cv_resnet18_ocr-detection-line-level_damo，这么小的模型为啥开销这么大，是有什么config参数没改吗？ 除了ocr-detection, 其他模块模型开销都在1s以内 我的环境是：cuda11.8, cudnn8.6.0, tensorflow2.11.0

如题： Application/DocXChain中的TableParsing，有现成工具，将其识别结果转为latex格式吗？