Lei Yang

@juxuan27 Could you please help update the latest progress here?

@akk-123 Hi, basically there are two ways. (1) You may try pass constructed smplx `verts` to [`visualize_smpl_pose`](https://github.com/open-mmlab/mmhuman3d/blob/main/mmhuman3d/core/visualization/visualize_smpl.py#L1152) directly; (2) Or you may pass `poses` dict to `visualize_smpl_pose`, where `poses` contains...

@jessica-santos change `.data` to `.data.item()` may work for you

[Paper](https://openreview.net/pdf?id=HJWLfGWRb) 第四页有详细说明，这篇[博客](https://jhui.github.io/2017/11/14/Matrix-Capsules-with-EM-routing-Capsule-Network/)中也有一些介绍。我对这个领域不熟悉，我的理解是这样的：作者想要让相同类型的capsule处理相同类型的entity，但同时又想要保留位置的信息，使得不同位置output的值不同，所以把坐标加进vote matrix。 我觉得也可以类比于conv来理解，比如一张image上有4个车轮，那这4个车轮可以用同样的conv来提取特征，这是卷积核的局部共享。capsule在最后一层的时候，既想要有这种共享的transformation，同时也想encode位置信息，所以采用了Coordinate Addition的技术。作者在Table 1的实验表明（对比Row.3和Row.6），使用这个技术的效果是有提升的。 这是我的理解，欢迎讨论。

Thanks for the discussion. There are also other issues or blogs concerning this question: - https://github.com/naturomics/CapsLayer/issues/30 - https://yhyu13.github.io/2017/12/17/Matrix-Capsule-With-EM-Routing-Reproduce-Report/ Reproduce the CNN baseline maybe a good start point to explore this...

@tomahawk810 Sorry about the late reply, as I am busy preparing cvpr submission. As stated in https://github.com/yl-1993/Matrix-Capsules-EM-PyTorch/blob/9cb3fc04442b937a1769d6d62e86d1754b7d9083/model/capsules.py#L192 The shape of `x` after adding patches is (b, H', W', K, K,...

@tomahawk810 I get your point now. I check the code and you are right about the output shape. The issue lies in the `add_patches` function. I have fixed the problem...

@tomahawk810 I merge the PR since there is no comment for two weeks. Please continue to comment if you feel anything is weird.

Both of them work. https://github.com/deepinsight/insightface is also great. The key is to train a face recognition model to extract face features.

Have you tried after this PR https://github.com/yl-1993/learn-to-cluster/pull/44? It seems to remove the need for .meta during inference. Let me know if there is any left issue.