stereo-dataset/stereo-dataset-small-sample-2gb

Stereo Dataset Small Sample

This repository is a size-constrained sample drawn from the full Stereo Dataset release.

Source dataset

• Full dataset: stereo-dataset/stereo-dataset
• Source repo commit at refresh time: 46d28a23024387da97516b160cce35c58ffa2e60
• Only complete scenes containing _scene_complete.json were eligible for sampling.

This 16-scene sample is provided so users can inspect dataset structure and quality without downloading the full 1.47 TB release. The full dataset contains 8,493 scenes (7,754 train, 739 eval); see the link above.

Scene files and geometry contract

The sample preserves the original scene directory structure and file names from the full dataset. Each selected scene_XXXXXX/ directory contains:

File	Purpose
_scene_complete.json	scene completeness marker
baseline.json	realized adjacent/pairwise baselines + camera intrinsics
trajectory.json	per-frame 6-camera poses
cam_00_rgb.mp4 ... cam_05_rgb.mp4	synchronized RGB videos
cam_00_depth.mkv ... cam_05_depth.mkv	log-encoded metric depth videos

RGB, depth, and pose records are aligned by frame index: frame t in cam_XX_rgb.mp4 corresponds to frame t in cam_XX_depth.mkv and frames[t] in trajectory.json. Use this frame-index contract for alignment; do not infer RGB/depth sync from container timestamps alone. RGB MP4 streams are 1280 x 1280, 15 fps, 81 frames. Depth MKV streams have the same spatial resolution and frame count, but the checked release files report 30 fps in the FFV1 stream metadata.

Depth videos are FFV1/gray16le uint16 videos with logarithmic metric-depth encoding. Decode stored uint16 values v to meters as:

import numpy as np


def decode_depth_uint16(v: np.ndarray) -> np.ndarray:
    v = np.asarray(v, dtype=np.uint16)
    depth_m = np.full(v.shape, np.nan, dtype=np.float32)

    valid = v > 0
    t = (v[valid].astype(np.float32) - 1.0) / 65534.0
    depth_m[valid] = 0.1 * (50000.0 ** t)
    return depth_m

v == 0 is invalid. v == 65535 is the maximum encoded value and decodes to 5000 m. It can include depths clipped to that upper bound, so evaluations that exclude far or saturated regions may choose to mask it.

Intrinsics are stored in baseline.json under camera_intrinsics. For image width W and height H, compute pixel focal lengths as:

fx_px = focal_length_mm / sensor_width_mm * W
fy_px = focal_length_mm / sensor_height_mm * H

If no principal-point override is recorded in baseline.json, use the image center for (cx, cy).

baseline.json stores realized scalar baseline distances in centimeters in adjacent_pairs and pairwise_pairs. Convert to meters before combining with decoded depth. The released rig is intended as a rigid rectified lateral camera array, so for a selected rectified pair:

disparity_px = fx_px * (baseline_cm / 100.0) / depth_m

This is a sanity check for a matched camera pair and frame. If it appears inconsistent with visible RGB disparity, verify the depth decoding, cm-to-meter conversion, pixel fx computation, camera-pair selection, pair direction, and RGB/depth frame alignment.

For full projective evaluation, use trajectory.json poses together with the intrinsics from baseline.json. trajectory.json stores camera centers in the Unreal/world coordinate frame in centimeters and rotations as quaternion_xyzw. These records are exported from the Unreal camera actor pose: the optical axis is the actor forward vector, image x uses the actor right vector, and image y uses the actor up vector. Compose source-to-target relative poses in one chosen coordinate convention; if your pixel coordinate system has y increasing downward, handle that sign convention explicitly.

Sampling procedure

The sample was created automatically from the finalized dataset snapshot using a reproducible random procedure:

1. Enumerate all complete scenes under train/ and eval/.
2. Shuffle the eligible scene list with random.Random(20260506).
3. Traverse the shuffled list once and greedily keep a scene only if adding it keeps the cumulative on-disk size under the target budget.
4. Materialize the selected subset while preserving each scene's original relative directory path.

This means the sample is random but size-constrained; it is not stratified by split, family, or map.

Sample summary

• Selected scenes: 16
• Total size: 1.800 GiB
• Target budget: 1.800 GiB
• Random seed: 20260506
• Target HF repo: stereo-dataset/stereo-dataset-small-sample-2gb

Sampling preserved the original directory structure under train/ and eval/.

Bucket counts

• eval/MapSeenInTrain/IPD_Gaussian/GangnyeongieonComplex: 1
• eval/MapSeenInTrain/Uniform/GeunjeongjeonComplex: 1
• train/IPD_Gaussian/GangnyeongieonComplex: 2
• train/IPD_Gaussian/GeunjeongjeonComplex: 1
• train/IPD_Gaussian/RestaurantScene: 1
• train/Pairwise_Uniform/AbandonedPowerPlant: 1
• train/Pairwise_Uniform/AssetsvilleTown: 3
• train/Pairwise_Uniform/GangnyeongieonComplex: 1
• train/Pairwise_Uniform/JesuhabComplex: 2
• train/Pairwise_Uniform/SeyeonjeongPavilion: 1
• train/Pairwise_Uniform/UndergroundSciFi: 1
• train/Uniform/ProceduralBuildingGenerator: 1

See sample_manifest.json for the exact scene list and size statistics.