search

ZhangMingyang-cs / Action-Recognition

用ResNet做基于骨架的动作识别
MIT License
6 stars 1 forks source link

tree_traversal_one_ske #2

Open sorrycy opened 1 year ago

sorrycy commented 1 year ago

你好，以下这段代码是什么意思啊，没太看懂 def tree_traversal_one_ske(ske_joint, dataset): shape = ske_joint.shape ret = np.zeros((shape[0], 1 + 2 * (shape[1] - 1), 3)) if dataset in ['NTU', 'NTU120']: sequence = [1, 20, 2, 3, 2, 20, 4, 5, 6, 7, 21, 7, 22, 7, 6, 5, 4, 20, 8, 9, 10, 11, 22, 11, 24, 11, 10, 9, 8, 20, 1, 0, 12, 13, 14, 15, 14, 13, 12, 0, 16, 17, 18, 19, 18, 17, 16, 0, 1]

ZhangMingyang-cs commented 1 year ago

类似于深度遍历，把各个关节串一遍，sequence是关节的序号

sorrycy commented 1 year ago

您好，我还是不太明白为什么转化成图之前要先进行树的遍历呢？，直接转换会丢失信息啥的吗？

ZhangMingyang-cs commented 1 year ago

您好，我还是不太明白为什么转化成图之前要先进行树的遍历呢？，直接转换会丢失信息啥的吗？

树的遍历之后，相邻的关节在图里就是相邻的点，ieee tpami有篇文章用了这个方法，实验的时候加上这个操作不一定能提高准确率，可以尝试一下

sorrycy commented 1 year ago

好的，我刚刚想起来在那儿看到过，应该是Spatio-Temporal LSTM with Trust Gates for 3D Human Action Recognition这篇文章，我去尝试一下 image

ZhangMingyang-cs commented 1 year ago

是这一篇

sorrycy commented 1 year ago

好的，非常感谢！

sorrycy commented 1 year ago

image 您好，为啥准确率能到100，loss还是6点多呀，不应该是0吗