tarashakhurana / 4d-occ-forecasting

CVPR 2023: Official code for `Point Cloud Forecasting as a Proxy for 4D Occupancy Forecasting'
https://www.cs.cmu.edu/~tkhurana/ff4d/index.html
MIT License
210 stars 22 forks source link

Can we extend when input resolution(voxel_size) is different in xyz dimensions? #13

Closed synsin0 closed 6 months ago

synsin0 commented 7 months ago

I see both gt_dist and pred_dist are multiplied with voxel_size after, what if voxel size is different on xyz dimensions?

tarashakhurana commented 6 months ago

You can have different resolutions across all three axes!

First of all, the cuda kernel is independent of the voxel resolutions. It only thinks of a grid as cells with integer-valued indices.

Second, the multiplication of gt_dist with voxel_size of 0.2 is actually a multiplication with $\frac{\sqrt{0.2^2 + 0.2^2 + 0.2^2}}{\sqrt{1^2 + 1^2 +1^2}}$. You could instead have different resolutions with different axes and multiply gt_dist and pred_dist with $\frac{\sqrt{v_x^2 + v_y^2 + v_z^2}}{\sqrt{1^2 + 1^2 +1^2}}$. Let me know if you try this!