Yu Sun

Results 301 comments of Yu Sun

Set it at https://github.com/Arthur151/ROMP/blob/653a0c9de13c7e242bc304147ae6559d1c6ff283/configs/webcam.yml#L36 It incluence the code at https://github.com/Arthur151/ROMP/blob/653a0c9de13c7e242bc304147ae6559d1c6ff283/romp/lib/visualization/create_meshes.py#L51

Those are texture UV maps of SMPL mesh. Some of those are obtained from https://github.com/bharat-b7/MultiGarmentNetwork#dress-smpl-body-model-with-our-digital-wardrobe

Please set the config setting show_largest: False Besides, I think it is highly feasible.

您好, 这种单阶段的网络训起来确实是比较慢,它的检测能力和人体姿态恢复能力是交替提升的。 我看了下你的训练log,看起来没什么问题。 如果可以的话,还是建议用最新的训练代码,做了一些必要的优化,并且之前训练存在的问题也解决了一些。可能validation时有些bug,但是训练效果会更好些。 我没有刻意的留存这些log文件,抱歉。之前有些issue里好像有其他人训练的记录。

1.训练loss的反复跳动是正常现象。 2.那个bug是有关于pytorch3D渲染的,和训练无关。 I recommand that you can go with the protocols using 3DPW training set during training, and evaluate on 3DPW test set. Protocols training without 3DPW is counting on...

3DPW的测试集和验证集是在这里加载: https://github.com/Arthur151/ROMP/blob/91dac0172c4dc0685b97f96eda9a3a53c626da47/romp/base.py#L170 在这里使用: https://github.com/Arthur151/ROMP/blob/91dac0172c4dc0685b97f96eda9a3a53c626da47/romp/train.py#L117 https://github.com/Arthur151/ROMP/blob/91dac0172c4dc0685b97f96eda9a3a53c626da47/romp/train.py#L142 不会用于训练。 训练集是通过这里加载: https://github.com/Arthur151/ROMP/blob/91dac0172c4dc0685b97f96eda9a3a53c626da47/romp/base.py#L126 也就是说只是用了3DPW的训练集训练。用验证集验证,然后选在验证集上表现最好的checkpoint来测试集评测。 使用3DPW的训练集训练可以帮助模型跨过数据集之间的domain gap(关节点定义不同,相机拍摄条件不同等等),所以会表现更好。

是的。两种评测protocol,不同的表格进行对比。 在3DPW上,完全可以直接和加入pw3d之后训练出来的模型,进行结果对比。

Hi, Vivian, Thanks for reporting this bug. I will take care of it.

Great idea. I have been working on this problem too. Following the way that CHD explored, I am trying to develop a more concise/compact method to achieve the dynamics smoothness,...

@visonpon ,Hi, I am trying to connect ROMP and contact_human_dynamics. Could you please provide the output file of contact_human_dynamics, after running: python run_totalcap.py --data ../data/example_data --out ../output/mtc_viz_out --totalcap ../external/MonocularTotalCapture Thanks...