Dynamic Object Motion Disentanglement (DOMD)

[757787182]

I'm sorry to bother you,but I have a question that has puzzled me.the three picture in DynamicDepth Ct-1 or Ct-1d,which one is the the true in frame t-1which one is crop,and what is the purpose of this，It's hard for me to understand the paper and the three picture.

[fengziyue]

Hi:

Thank you so much for your interest in our work!

Do you mean the Fig.2 in the main paper? $I_{t-1}$ is the true frame t-1 while the $I^d_{t-1}$ is the dynamic object motion disentangled image or you called cropped one. It shows the object at time t's (real world) 3D location but in t-1's camera perspective.

So the $I^d_{t-1}$ and $I_t$ shows different perspective of the object at same 3D location, as if it's a static object. to simplify the downstream models work.

I hope this solves your question, let me know if you have any further concerns!

[757787182]

thank you for your reply. I have figured it out，but I have another question and I hope you can help me.，How do you estimate occlusion issues between t-1 and t+1 frames in Occlusion-aware Re-projection Loss Lor? and How to determine occlusion for dynamic objects ？

[fengziyue]

Sorry for the late reply, since we have the segmentation mask for all dynamic objects, we can easily get the occlusion mask via re-projection.