Infinite depth issue

[ahtso]

Hi @Lavreniuk

Last time I asked you about the backbone issue of PoseNet #4
Now I have successfully trained,
but I found that my dataset (non KITTI or Cityscapes) will have serious illness problem of infinite depth.
So I have re-trained with the filtered dataset by removing the frame by frame pixel difference is too small.
The improvement is limited.
Could you please give me some advice how you solved this kind of problem in the past, besides AutoMask?

[Lavreniuk]

Hi @ahtso, glad to hear that you've managed to get the training running. When it comes to handling the issue of infinite depth, automasking has been the primary approach I've used. Other than automasking, I haven't employed additional tricks. However, you might consider a few options to further address this problem:
Depth Range Clipping: Manually set a maximum depth threshold to prevent extremely large depth values. This can help avoid the impact of outliers.
Normalization or Scaling of Depth Values: Ensure that the depth values are normalized or scaled appropriately to maintain a stable training process.

[Lloyd0310]

你好

上次我问你关于 PoseNet #4 的 backbone 问题 现在我已经训练成功了，但是我发现我的数据集（非 KITTI 或 Cityscapes）会出现无限深度的严重疾病问题。所以我用过滤后的数据集重新训练，去除了太小的逐帧像素差异。改进是有限的。您能否给我一些建议，除了 AutoMask 之外，您过去是如何解决此类问题的？

你好，我想知道你是否已经解决无限深度的问题，因为我同样遇到了该问题，想寻找好的方法解决。
Hi, I would like to know if you have resolved the infinite depth issue as I am also experiencing the issue and would like to find a good way to resolve it.