cwpeng

I noticed that many papers mentioned that the accuracy of image-based training is lower than that based on time series average pooling.  But I found that in your project,...

您好，我使用了您提供的basenet.pth，运行test之后能够在DUT-TE上得到与论文一致的精度。 之后我尝试在train中把basenet.pth直接load进来，进行finetune。训练数据集仍然为DUT-TR，代码无改动。 但非常奇怪的是，发现初使的loss非常高，如下所示，这是否不合理？因为这个trainl oss明显还能进一步下降，远远没有拟合训练集。按论文描述，迭代400k步后，那loss应该很小。请问我有弄错什么吗？ [epoch: 1/100000, batch: 32/10553, ite: 4] train loss: 8.034277, tar: 0.670928 l0: 0.826161, l1: 0.863950, l2: 0.952028, l3: 0.988486, l4: 1.222528, l5: 1.460253, l6: 1.663608

Here are three functions related to data augmentation，include pose_processing，j2d_processing and j3d_processing. [Code Link ](https://github.com/nkolot/SPIN/blob/master/datasets/base_dataset.py) When doing rotation augmentation on 2d joint data (j2d_processing), the rotation is around the center of...