yiranranranra/stereo-vipe-eval

Stereo ViPE evaluation clips

The eight calibrated stereo clips used to evaluate Stereo ViPE, the calibrated-stereo version of NVIDIA ViPE. Each clip comes with its left and right images, ground-truth camera poses, ground-truth depth and calibration, plus the LinStereo depth and disparity computed for every frame and both views. The TartanAir and KITTI archives keep their original layout; extracted as below, they sit where the Stereo ViPE scripts read them. Stereo ViPE reads the LinStereo archives directly (stereo_cli infer --precomputed), so nobody needs to run the stereo network.

Clip	Source	Frames	Images	Ground truth	Archive
P000	TartanAir v2, ArchVizTinyHouseDay, easy	116	640×640 RGB, left + right	poses, dense depth (left + right)	tartanair/ArchVizTinyHouseDay_Data_easy_P000.tar.zst
P001	〃	120	〃	〃	…_P001.tar.zst
P002	〃	134	〃	〃	…_P002.tar.zst
P003	〃	152	〃	〃	…_P003.tar.zst
P004	〃	135	〃	〃	…_P004.tar.zst
P005	〃	118	〃	〃	…_P005.tar.zst
P006	〃	91	〃	〃	…_P006.tar.zst
KITTI 04	KITTI odometry sequence 04	271	1226×370 RGB, image_2 + image_3	poses (poses/04.txt)	kitti/odometry_04.tar.zst
KITTI 04 depth	KITTI depth, annotated, drive 2011_09_30_drive_0016 (val)	269	—	LiDAR depth for image_02 + image_03	kitti/depth_annotated_2011_09_30_drive_0016.tar.zst

linstereo/linstereo_<clip>.tar.zst (P000 … P006, K04) holds the LinStereo output for every frame and both views: disparity, depth and a validity mask, with the intrinsics and baseline (see LinStereo depth and disparity).

previews/*.mp4 shows each clip as left | right at half resolution (H.264, 10 fps) for quick browsing; use the archives for evaluation. SHA256SUMS lists every file.

In the Stereo ViPE experiments P000 is the development sequence; P001–P006 are held-out trajectories in the same scene, with one exception: the depth-fixed-plus-flow ablation was reworked after its first implementation diverged, its discarded P001–P003 results were seen, and the first 24 frames of P001 were used to diagnose the divergence. KITTI 04 had been used in earlier development diagnostics, so it is not held-out.

Download and extract

hf download yiranranranra/stereo-vipe-eval --repo-type dataset --local-dir stereo-vipe-eval
cd stereo-vipe-eval && sha256sum -c SHA256SUMS

# the directory names the Stereo ViPE scripts read (under <project>/data); the KITTI poses also go
# to a separate ground-truth directory, which the report and video scripts read
D=<project>/data
mkdir -p $D/tartanair-v2-archviz-tiny-house-day-easy-208ede3 $D/kitti-odometry-official \
         $D/kitti-odometry-ground-truth-20260929T042443Z $D/kitti-depth-annotated-2011_09_30_drive_0016-20261011
for f in tartanair/*.tar.zst; do tar --zstd -xf "$f" -C $D/tartanair-v2-archviz-tiny-house-day-easy-208ede3; done
tar --zstd -xf kitti/odometry_04.tar.zst -C $D/kitti-odometry-official
tar --zstd -xf kitti/odometry_04.tar.zst -C $D/kitti-odometry-ground-truth-20260929T042443Z dataset/poses/04.txt
tar --zstd -xf kitti/depth_annotated_2011_09_30_drive_0016.tar.zst -C $D/kitti-depth-annotated-2011_09_30_drive_0016-20261011

# LinStereo output: one directory per clip (P000 … P006, K04)
mkdir -p linstereo-out && for f in linstereo/*.tar.zst; do tar --zstd -xf "$f" -C linstereo-out; done

tar --zstd needs GNU tar ≥ 1.31; otherwise use zstd -dc FILE | tar -xf -.

Layout and formats

TartanAir v2 (P000–P006)

ArchVizTinyHouseDay/Data_easy/P000/
  image_lcam_front/000000_lcam_front.png …        left RGB, 640×640
  image_rcam_front/000000_rcam_front.png …        right RGB
  depth_lcam_front/000000_lcam_front_depth.png …  left depth
  depth_rcam_front/000000_rcam_front_depth.png …  right depth
  pose_lcam_front.txt, pose_rcam_front.txt        one line per frame

• Calibration (TartanAir v2 front pinhole pair, as used by Stereo ViPE): fx = fy = 320, cx = cy = 320 (pixels), 640×640, rectified, baseline 0.25 m, no distortion.
• Depth is metric z-depth stored as an RGBA uint8 PNG whose four bytes are one little-endian float32. Decode with cv2.imread(path, cv2.IMREAD_UNCHANGED).view('<f4')[..., 0] (do not reorder channels). Far background reaches very large values, so Stereo ViPE scores depth only in a range: its evaluator and the report's per-sequence depth table use true depth 0.1–100 m; its README depth table, the comparison with the rig control, and the slides and videos use ≤ 50 m.
• Poses: tx ty tz qx qy qz qw, camera-to-world, in TartanAir's NED convention (camera axes x forward, y right, z down). Convert to OpenCV camera axes (x right, y down, z forward) before comparing with an OpenCV-convention estimate.
• Time: TartanAir v2 frames carry no timestamps; Stereo ViPE uses the frame index at a nominal 10 Hz.

KITTI odometry 04

dataset/sequences/04/image_2/000000.png … 000270.png   left colour camera, 1226×370
dataset/sequences/04/image_3/000000.png … 000270.png   right colour camera
dataset/sequences/04/calib.txt                         P0–P3 (3×4), Tr
dataset/sequences/04/times.txt                         seconds per frame
dataset/poses/04.txt                                   3×4 camera-to-world, row-major, one line per frame

Images are KITTI's rectified colour images. The ground-truth poses are those of camera 0 (left grey camera). The left colour camera image_2 sits at the offset given by K⁻¹ P2 from calib.txt (about 6 cm along x); Stereo ViPE's evaluator applies it.

KITTI annotated depth for odometry 04

val/2011_09_30_drive_0016_sync/proj_depth/groundtruth/image_02/0000000005.png … 0000000273.png
val/2011_09_30_drive_0016_sync/proj_depth/groundtruth/image_03/…

• Raw drive 2011_09_30_drive_0016 is odometry sequence 04 with the same frame numbers: raw frame i = odometry frame i. The annotations cover raw frames 5–273, so odometry frames 5–270 have depth. Stereo ViPE checked the mapping: its depth error (AbsRel) against this ground truth is 2.0 % at offset 0 and 4.2–4.3 % at ±1 frame.
• Format: uint16 PNG, 1226×370; depth in metres = value / 256; 0 = no measurement.

LinStereo depth and disparity

P001/release.json                    binds each disparity file (sha256) to its frame, for the pipeline
P001/meta.json                       sequence, K, baseline, image size, field definitions, valid fraction
P001/geometry.json                   rig geometry Stereo ViPE used (camera-to-reference, metres)
P001/left/000000_disparity_px.npy    float32 H×W
P001/left/000000_depth_m.npy         float32 H×W
P001/left/000000_valid.npy           bool    H×W
P001/right/…                         the same for the right view
K04/rectified/left/000000.png …      KITTI only: the rectified images the depth refers to

• LinStereo runs on every frame at full resolution, in both directions (left→right and right→left). The file prefix is the frame index of the source clip.
• disparity_px is the disparity magnitude in pixels, positive in both views: the left-view match is at x − d in the right image, and the right-view match at x + d in the left image. It is the raw output at every pixel, so use it only where valid.
• depth_m is camera z-depth in metres, fx · baseline_m / disparity_px where valid and 0 elsewhere.
• valid marks finite, positive disparity whose match lands inside the other image and passes a 1-pixel left–right consistency check.
• release.json records, for every frame, the physical pair key (a hash of the source bytes and the preparation), the grid ids and each disparity file's sha256.
  • stereo_vipe.stereo_cli infer --precomputed <dir> checks all three and recomputes valid, depth and inverse depth from the disparity with the pipeline's own code.
  • This gives exactly the arrays Stereo ViPE used. A different preparation or a changed file stops the run.
• TartanAir: pixel-aligned with image_lcam_front / image_rcam_front. K is fx = fy = 320, cx = cy = 320, and the baseline is 0.25 m.
• KITTI 04: pixel-aligned with K04/rectified/{left,right}. These are image_2 / image_3 re-rectified by a small pure rotation Rrect (0.47°, in geometry.json) onto a common K (fx = 725.17) and padded to 1232×376; the baseline is 0.5355 m. meta.json gives the exact mapping back to the original KITTI pixels.

Licences

The dataset archives (tartanair/, kitti/) are redistributed unchanged (only repacked) under their original licences. The preview videos are re-encoded copies. The LinStereo files are model outputs computed from these images, and linstereo_K04 also contains re-rectified KITTI images. The LinStereo files are for non-commercial research use and also carry the licence of the images they come from. See LICENSE.md.

• TartanAir v2: CC BY 4.0, © the TartanAir authors (tartanair.org).
• KITTI odometry and KITTI depth: CC BY-NC-SA 3.0, © Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago (cvlibs.net/datasets/kitti). Non-commercial use only.

Citation

If you use these clips, cite the original datasets:

@inproceedings{wang2020tartanair,
  title     = {TartanAir: A Dataset to Push the Limits of Visual SLAM},
  author    = {Wang, Wenshan and Zhu, Delong and Wang, Xiangwei and Hu, Yaoyu and Qiu, Yuheng and
               Wang, Chen and Hu, Yafei and Kapoor, Ashish and Scherer, Sebastian},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2020}
}
@inproceedings{geiger2012kitti,
  title     = {Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite},
  author    = {Geiger, Andreas and Lenz, Philip and Urtasun, Raquel},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2012}
}
@inproceedings{uhrig2017sparsity,
  title     = {Sparsity Invariant CNNs},
  author    = {Uhrig, Jonas and Schneider, Nick and Schneider, Lukas and Franke, Uwe and Brox, Thomas
               and Geiger, Andreas},
  booktitle = {International Conference on 3D Vision (3DV)},
  year      = {2017}
}