search

Lilac-Lee / FastNSF

Fast Neural Scene Flow (ICCV 2023)
https://Lilac-Lee.github.io/FastNSF
MIT License
71 stars 4 forks source link

performance for moving objects #4

Open waveleaf27 opened 1 year ago

waveleaf27 commented 1 year ago

I re-ran this project, but the performance for moving objects didn't meet expectations. Please refer to the attached figure for reference.

In the figure: (segment-10335539493577748957_1372_870_1392_870_with_camera_labels/0000_0001.npz) image

Black represents points at time T. Blue indicates points at time T+1. Red denotes points from time T after adding the predicted flow. As evident, the objects within the bounding box are misaligned. I've observed this issue in multiple case. Is this discrepancy expected?

@Lilac-Lee

waveleaf27 commented 1 year ago

Can you provide your NSFP++ baseline since it adopt extra assumption

Lilac-Lee commented 1 year ago

Hi @waveleaf27, this might happen when the grid cell size is relatively large, making the prediction prone to similar motions nearby. Please refer to MBNSF for the implementation of NSFP++ baseline. Cheers.

Kin-Zhang commented 3 months ago

I also observed the moving object prediction in FastNSF is not better than NSFP in AV2 too. But in general, it's not bad.

Btw, I want to ask about grid cell size is that equal to grid_factor in the code? Is 10 default mean 0.1-meter grid size?