About Depth to Disparity

[cocowy1]

Hello author,
I have some questions regarding the depth-to-disparity conversion in the depth_splat.py script. The standard formula for converting depth to disparity is typically given as:

$disparity= \frac{depth}{baseline×focal length}$

However, in the implementation, it appears that the camera's internal parameters (such as focal length and baseline) are not explicitly defined.

disp_map = torch.from_numpy(batch_depth).unsqueeze(1).float().cuda()
disp_map = disp_map * 2.0 - 1.0
disp_map = disp_map * max_disp

Could you clarify why these parameters are not explicitly included in the code? Is it because they are implicitly handled elsewhere, or are there assumptions about the camera setup that allow for this simplification?

Additionally, I noticed that the weight map in the depth splitting function is computed as:

weights_map=(1.414) ** weights_map

This operation seems to scale the weight map by a factor of $\sqrt{2}$, but there is no mention of why using this in the paper. Could you explain the rationale behind this specific scaling factor? Is it related to normalization, optimization, or some other aspect of the algorithm?

[hengshan123]

I have same questions

[ZitongLan]

Yes, I have the same question, dear author @xiaoyu258 can you help us with this problem?

[ZitongLan]

Also, from the a quick google search, it seems that the disparity is inversely proortional to the depth, but here it seems it is positively related, what is wrong?

[ZitongLan]

I dig into the code and found that the depth is first normalized into 0 to 1, where 0 is farther and 1 is closer. Then it satisis this negative proportional relationship.