Datasets:
Clarification on depth decoding
Thank you for releasing StereoGenBench.
I am trying to use stereo-dataset-small-sample-2gb for stereo/depth evaluation, but the RGB stereo disparity does not seem to match the disparity implied by the released depth and camera metadata.
I checked the standard stereo relation, disparity = fx * baseline / depth, using baseline.json, the pairwise baseline, and several plausible decoding conventions for *_depth.mkv. However, the visible RGB disparity still appears inconsistent with the decoded depth.
Interestingly, projecting between the released depth maps using the recorded poses gives good depth-depth consistency. This makes me wonder if I am missing a depth decoding rule, camera convention, or RGB/depth synchronization detail.
Could you clarify the intended decoding/evaluation contract for the *_depth.mkv files?
Thanks for the careful report and for checking the depth-depth consistency with the released poses. You are right that the current README does not spell out the full user-facing RGB/depth/camera evaluation contract clearly enough. Some of the detailed geometry checks are already in our manuscript/evaluation notes, but the public dataset README should also contain a concise contract, and I will make justification ASAP. The main work and detailed explanation has been included in paper, but out of double-blind consideration, it is better not to be released it in the page.
For the released *_depth.mkv files, the intended decoding is logarithmic metric depth, not a linear uint16 scale. In the production conversion path, valid values use:
valid = (v > 0)
t = (v.astype("float32") - 1.0) / 65534.0
depth_m = 0.1 * (50000.0 ** t)
depth_m[~valid] = np.nan
Here v == 0 is reserved for invalid pixels. Values near the upper end correspond to clipped/far-background depths; if you want to exclude sky/saturated regions, masking v == 65535 is reasonable.
On the stereo/disparity side, the released rig is intended to be a rigid rectified lateral camera array. Therefore, for a selected rectified pair, disparity = fx * B / z should be the expected sanity check.
If the RGB-visible disparity appears inconsistent with that value, the first things I would verify are the depth decoding above, cm-vs-meter unit conversion, pixel fx computation, exact camera-pair selection / pair direction, and RGB-depth frame alignment. The per-frame poses in trajectory.json, together with the intrinsics, are the authoritative way to reproduce the projective geometry and debug any remaining mismatch.
I agree this should be documented better. I am going to update the full dataset README and the small-sample README with a dedicated depth/camera-geometry section covering depth decoding, units, invalid/clipped values, intrinsics-to-pixel fx/fy, pose convention, frame alignment, and the relationship between baseline.json, trajectory.json, and rectified-pair disparity.
Could you share one concrete scene path, camera pair, and frame index where the RGB disparity mismatch is most visible? I can use that as a validation case while updating the documentation.