Datasets:

---

fpvlabs

/

stereo-550

Like 2

Follow

FPV Labs 70

Hi FPV Labs team — thank you for releasing Stereo-550 and the Ego-OSCAR paper (arXiv:2608.08285).

While validating metric depth, we found a reproducible scale inconsistency between the packaged stereo videos and calibration.json. We are not claiming that the physical 42 mm baseline is wrong; our working hypothesis is that a video split/resize/crop convention may be missing from the released metadata or preprocessing instructions.

Reproduction

We tested the following original Stereo-550 sessions:

stera001_0dfn9sa0v5tw
stera002_0y5uzrjnxa9a
stera001_6bf5y5svcg1u

We used the supplied 1280×720 left.mp4, right.mp4, camera intrinsics, distortion, stereo rotation/translation, and baseline_mm.

Images were rectified with OpenCV stereoRectify using CALIB_ZERO_DISPARITY. For stera001, the rectified focal length is approximately 498.313 px and the supplied baseline is 42.137958 mm.

Depth was computed as:

Z_m = fx_rect * (baseline_mm / 1000.0) / disparity_px
No additional scale factor was used for the measurements below.

Three independent disparity estimates agree

At the hand keypoints in frame 0 of stera001_6bf5y5svcg1u:

OpenCV SGBM: median disparity 120.8 px
Fast-FoundationStereo: median disparity 121.4 px
Independent RTMPose detections on left/right images: median disparity 120.6 px
The methods agree within 1 px, so this does not appear to be a FoundationStereo failure.

The left/right keypoints also have small vertical residuals after rectification, indicating that they correspond to the same physical hand.

Metric-scale inconsistency

For stera001_6bf5y5svcg1u, wrist-to-middle-MCP spans approximately 79 px.

Using the supplied baseline and measured disparity:

79 px × 42.137958 mm / 121 px ≈ 27.5 mm
An adult wrist-to-middle-MCP distance is normally around 80–100 mm.

The same check across different episodes and devices produced:

stera001_6bf5y5svcg1u: approximately 27.3 mm
stera001_0dfn9sa0v5tw: approximately 25.1 mm
stera002_0y5uzrjnxa9a: approximately 23.2 mm
The supplied baseline also puts the head-mounted camera only about 0.2 m above the fitted floor plane in the first episode. Dense depth for frame 0 is approximately 0.11–0.24 m, although the paper describes a typical reliable depth range of 0.5–4 m.

Multiplying metric scale by roughly 3.5–4 gives plausible hand dimensions and camera-to-floor distances. We do not think the physical module baseline is actually 150–170 mm; this looks more like an unresolved calibration/video-coordinate scale convention.

Relative geometry appears correct

The reconstructed floor is strongly planar. Under the supplied scale, its plane-fit RMS is approximately 2.1 mm.

We also tested subtracting a constant disparity offset. This progressively warped the floor plane, so the issue does not appear to be an additive disparity bias. The disparity field appears geometrically coherent but globally mis-scaled.

Calibration-file observation
In our downloaded subset:

141 stera001 episodes contain byte-identical cali.json files
26 stera002 episodes contain another byte-identical cali.json
The paper states that calibration is per-session. Could you confirm whether the released calibration is actually per-device?

Questions

Are the released 1280×720 per-camera videos in exactly the same pixel coordinate system used for calibration?
Was the original 2560×720 side-by-side stream resized, cropped, or split in a way that requires scaling disparity or projection parameters?
Should baseline_mm be used directly with disparity from the released MP4 files?
Is the provided disparity_to_depth matrix intended to output millimetres directly?
Are calibration files intentionally shared per device rather than generated per session?
Could you provide expected SGBM or RAFT-Stereo disparity/depth statistics for frame 0 of stera001_6bf5y5svcg1u?
We can provide scripts and visualizations if useful. Thanks for any clarification.