Shiyu Zhao

Hi @racinmat , thanks for your sharing. According to your thesis, it seems that the transformation from camera coordinate to NDC is perspective projection. So, does it matter that P_{3,2}...

Hey, we didn't evaluate the convergence rate, but you may reduce the update iterations by setting `iters` for [this function](https://github.com/xiaofeng94/GMFlowNet/blob/master/core/gmflownet_model.py#L86). `iters` should work even you don't train with a specific...

Hey, can you share more details about the results and the setting? BTW, you need to use the newest code to reproduce the results. I used a smaller learning rate...

Hey, can you reproduce the results with released weights? What's your environment? I'm wondering if you got different environments, like pytorch, cuda. New version of pytorch may change some default...

Hey, we also found the training was not quite stable on KITTI possibly because there are less data. Also, it seems HD1K data causes the instability. You may remove it...

Hey, We didn't save optimizer's state_dict. The `total_steps` is attained from the file name (usually it ends with the number of iterations). So when resume training, you may expect some...

Hey, why not take a try. I think `initial_lr` may be set somewhere in RAFT's codebase.

Hey, we only report the 2-view results. You may check the caption of Table 2 in the paper or run the experiments on your own.

Hey, 非常感谢关注，以下是我们的回答： 1. 我们并没有保存最优权重。并且不建议根据测试结果挑选最优权重，那可能会导致数据泄露和过拟合 2. 我们的图像融合方法是一个离线过程，不需要和特定的训练过程绑定

Hey, 我们总是使用训练结束得到的最后一个模型权重进行测试。训练iterations一般detectron2默认的90k。如果您想选择更优的权重，可以考虑使用验证集进行选择。