technical discussion

[yushengjiexy]

technical discussion

[gengshan-y]

There are a few possibilities:

1. Directly run banmo on videos of different cars. This can model category shapes by LBS-based deformations. But it is still slow.
2. Train a feed-forward network with ground-truth from 1. This will give you a fast inference pipeline.

UCMR did something similar to what I suggested above.
ViSER and DOVE combines 1 and 2 in a joint optimization pipeline.

[yushengjiexy]

Hello, I found that if only part of the cat can be seen in the video, the reconstructed model is incomplete.
Hi, if I want to model a car, but the car is only partially visible in the video, what is the recommended way to get the full car model? thanks

[gengshan-y]

One thing to try is the symmetry option. This is equivalent to augmenting the data to be mirror symmetric.

Another way is to jointly reconstruct multiple videos of similar-looking cars, while sharing the the same canonical shape.

[yushengjiexy]

I searched some single view reconstruction papers, such as pixelNeRF, but I didn't find the reconstruction of moving objects Occluded by multiple views. Do you have any papers recommended for this? thank!

[yushengjiexy]

if you have time, could you give me some ideas?Thank you!

[gengshan-y]

I would still suggest reading UCMR/DOVE/ViSER, which reconstruct category shapes from images/videos. Another related one is panoptic neural fields.

[yushengjiexy]

thanks