You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Point4D training data

Gated internal mirror. Each dataset lives under a neutral code (data-NNN/). Clip shards follow the p4d-1 / p4d-1.2 WebDataset format (one tar member per frame and modality, <shard>.idx.json member offsets for range reads), with a per-dataset data-NNN/meta.tar holding dataset.json (splits, licenses/provenance, conventions, depth_semantics, quality, multiview), index.parquet (one row per clip) and the shard indices. The exception is data-013, which holds per-(sequence, camera) source units read by its own streaming loader.

Conventions shared by all datasets

Depth confidence: optional depth_confidence.TTTT.png (uint8, 255 = full weight, 0 where depth is 0) multiplies the per-pixel depth regression weight (loss_weights(status, confidence=...)); see dataset.json modalities.depth_confidence.rule / calibration.
Cameras: OpenCV camera frame (x right, y down, z forward); extrinsics are world-to-camera 4x4 per frame; intrinsics 3x3 per frame (full K, skew allowed). Non-pinhole clips (p4d-1.2) add camera_model + per-frame camera_params (opencv radial-tangential, opencv_fisheye, aria_fisheye624); use p4d_data.camera for project / unproject / raymap.
RGB: source JPEG/PNG bytes where the source ships image files; otherwise lossless WebP (decoded frames or re-encoded PNG; pixels identical). Lossless WebP decodes at ~8 ms per 960x540 frame on one core (PIL / libwebp; PNG ~7 ms, JPEG ~3 ms): a 128-frame clip costs ~1 core-second, so give the loader enough decode workers.
Depth: metres, planar camera z unless meta.modalities.depth.convention == "ray" (fisheye). 0 = no value. depth_status (uint8 per frame; low 4 bits = class, bit 7 = heuristic label) says how to supervise every pixel: 0 valid, 1 sky (supervise as infinite depth / inverse depth 0), 2 far_bounded (one-sided loss: prediction must be

= the per-clip far_lower_bound_m), 3 transparent (follow the dataset's transparent convention below), 4 volumetric_opaque (no target), 5 volumetric_partial (weight by transmittance), 6 edge_uncertain (real data only; low weight), 7 hole (no target). Reference weights: p4d_data.depth_status.loss_weights.
Transparent conventions: first_surface (depth on the glass), last_opaque (depth sees through glass), sensor / estimated (unreliable on glass). Multi-layer depth (depth_layers): layer 1 = standard depth, deeper layers along the ray, one per media transition.
Tracks (tracks.npz): world xyz [T, N, 3], valid, visible, query frame/uv, instance ids where available. Static tracks are fixed world points sampled from depth and visible by depth test. Optional per-observation depth_residual (signed, metres) to downweight points inside objects.
Flow: channel_order "dx,dy" in pixels; forward = displacement of frame t's pixel to t+1, backward = to t-1; dataset.json["flow_convention"]; p4d_data.flow.standard_flow(clip) returns the standard for any clip.
Time: dataset.json["timing"] gives fps where the source documents it; otherwise time is the frame index (unknown fps) or the clips are unordered views / single images.
Multi-camera: dataset.json["multiview"]: single_camera, native (views inside each clip, per-view visibility) or per_camera_clips with sync_groups. p4d_data.multiview.per_camera_visibility returns visibility of any track in every synchronized camera.
Unordered views: clips with extra_meta.unordered_views = true are view sets, not smooth video: skip temporal smoothness / flow-style losses.
Camera motion: index.parquet has per-clip cam_* columns (definitions, thresholds and units in dataset.json["camera_motion"]). cam_motion_class: static / rotation_only (max angular parallax < 1 deg; rotation extent < 1 deg or >= 1 deg), translating clips slow / moderate / fast by the median per-frame image motion m = degrees(translation / median scene depth) + rotation (< 0.5, < 2, >= 2 deg/frame; ~10 px per deg at a 60 deg / 640 px view), unordered (view sets), single (one frame). Also: translation speed per frame relative to the median depth (and m/s for metric datasets with known fps), rotation deg/frame and deg/s (p50 / p90 / max), parallax (median baseline / depth at stride 8, max angular parallax), path length, total rotation, p90 accelerations, cam_jerk_score / cam_shaky (high-pass jitter: hand-held-like vs smooth). Multi-view clips: scalars describe the most-moving view, cam_motion_class_per_view lists every view. Speeds are per stored frame.
Quality: dataset.json["quality"] gives a 1-5 score and suggested weights per supervised quantity (multiply with the per-pixel depth_status weights).
Supervision tiers: dataset.json["quality"]["supervision_tier"] labels every dataset's depth as exact (rendered ground truth: the only supervision target, incl. fine detail / edges), sensor (laser / LiDAR / phone depth / digital twin: real but noisy at edges and thin structures), coarse (multi-view stereo: only coarsely correct; fine detail, foliage and sky boundaries are wrong), mixed (per clip, extra_meta.depth_kind) or none; and cameras as exact, sensor (device tracking), estimated (SfM / SLAM) or intrinsics_only. Intended use: train on exact data; use sensor / coarse depth and estimated cameras only as conditioning inputs ("guesses") for a synthetic-trained model that later fills in the real data. If coarse depth is supervised at all, do it at low resolution only (pooled, scale-invariant, dropping cells that contain sky / holes), never with gradient or edge losses.

Datasets

data-001

What: synthetic (rendered), dynamic humans/animals/objects. Structure: video clips.
Supervision tier: depth exact, cameras exact (supervision target).
Size: 2,161 clips, 269,748 frames, 324.4 GB; median 128 frames/clip; resolution 960x540; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, normals 100%, instance 100%, tracks 100%.
Tracks: ground-truth 3D trajectories (static + dynamic); visibility recomputed by depth test (max(3 %, 3 cm)); per-observation signed depth_residual stored (downweight points far inside surfaces).
Flow: none.
Depth: transparent surfaces = mixed (verified); mirrors unknown; volumetrics noisy; multi-layer: no.
Camera motion (index cam_*): static 67, rotation_only 126, slow 974, moderate 961, fast 33; median image motion 0.51 deg/frame; shaky 674 (31 %).
Quality score: 4/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dynamic_tracks': 1.0, 'dense_motion': 1.0, 'covisibility': 1.0}.
fog/smoke scenes: volumetric depth labels in depth_status
outdoor scenes: HDRI environment beyond the ground plane is far_bounded (>= farthest geometry)
outdoor scenes: source cameras were one frame late; corrected (camera of frame t+1), the last frame of each sequence is extrapolated (extra_meta.camera_extrapolated_frames)
some outdoor clips contain fast rendered ~360 deg camera spins (extra_meta.camera_events): avoid training pairs that span them

data-002

What: synthetic driving (game engine), moving traffic. Structure: video clips; stereo pairs (cam0/cam1) as separate clips linked by sync groups.
Supervision tier: depth exact, cameras exact (supervision target).
Size: 380 clips, 42,520 frames, 76.5 GB; median 128 frames/clip; resolution 1242x375; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, instance 100%, semantic 100%, flow 100%, scene_flow 100%, tracks 100%.
Tracks: static-scene tracks + dynamic tracks on moving vehicles (from per-frame vehicle poses; parked cars stay static).
Flow: forward/backward optical flow + scene flow stored.
Depth: transparent surfaces = first_surface (verified); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): static 20, slow 80, moderate 56, fast 224; median image motion 2.9 deg/frame; shaky 0 (0 %).
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dense_motion': 1.0, 'covisibility': 1.0}.
weather variants (fog/rain/...) share the clear-weather geometry; fog transmittance in depth_status
sky = semantic sky class (infinite depth target)

data-003

What: real imagery with depth rendered from multi-view-stereo meshes. Structure: unordered multi-view image sets (not video).
Supervision tier: depth coarse, cameras estimated (conditioning input / guess for a synthetic-trained model; exact quantities only where the tier says exact).
Size: 661 clips, 66,113 frames, 58.0 GB; median 128 frames/clip; resolution 768x576; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, tracks 100%.
Tracks: static tracks from depth + cameras.
Flow: none.
Depth: transparent surfaces = estimated (verified); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): unordered 661.
Quality score: 3/5; metric scale: False; suggested weights: {'depth': 0.6, 'camera': 0.8, 'static_tracks': 0.5, 'covisibility': 0.6}.
arbitrary scale (scale-invariant losses only)
mesh holes -> hole; depth edges -> edge_uncertain

data-004

What: synthetic (game engine) environments, mostly static. Structure: video clips (front camera).
Supervision tier: depth exact, cameras exact (supervision target).
Size: 1,531 clips, 181,849 frames, 215.1 GB; median 128 frames/clip; resolution 640x640; fps 10.0.
Modalities (share of clips): rgb 100%, depth 100%, flow 100%, tracks 100%.
Tracks: static tracks from depth + cameras.
Flow: optical flow stored (lossless).
Depth: transparent surfaces = last_opaque (verified); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): fast 1,531; median image motion 5.2 deg/frame; shaky 1,518 (99 %).
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dense_motion': 1.0, 'covisibility': 1.0}.
depth sees through glass (last_opaque)
sky via per-environment segmentation ids; invalid non-sky >= 1000 m (far_bounded)

data-005

What: synthetic (rendered), dynamic, same pipeline as data-001. Structure: video clips.
Supervision tier: depth exact, cameras exact (supervision target).
Size: 90 clips, 3,600 frames, 2.1 GB; median 40 frames/clip; resolution 960x540; fps 24.0.
Modalities (share of clips): rgb 100%, depth 100%, tracks 100%.
Tracks: ground-truth trajectories; visibility recomputed by depth test (max(3 %, 3 cm)) + depth_residual.
Flow: none.
Depth: transparent surfaces = mixed (documented); mirrors unknown; volumetrics outlier; multi-layer: no.
Camera motion (index cam_*): static 11, rotation_only 13, slow 31, moderate 31, fast 4; median image motion 0.45 deg/frame; shaky 23 (26 %).
Quality score: 4/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dynamic_tracks': 1.0, 'covisibility': 1.0}.
source outlier code (85 m) converted to invalid
smoke (coded as solid blocks in the source) -> hole (no target)
outdoor clips: cameras corrected by one frame as in data-001

data-006

What: synthetic (rendered) dynamic object scenes. Structure: video clips.
Supervision tier: depth exact, cameras exact (supervision target).
Size: 5,869 clips, 704,280 frames, 500.6 GB; median 120 frames/clip; resolution 512x512; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, instance 100%, tracks 100%.
Tracks: ground-truth 3D trajectories (dense sampling, static + dynamic).
Flow: none.
Depth: transparent surfaces = first_surface (documented); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): static 1,262, rotation_only 6, slow 2,386, moderate 2,213, fast 2; median image motion 0.34 deg/frame; shaky 0 (0 %).
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dynamic_tracks': 1.0, 'covisibility': 1.0}.
background is a dome mesh at 40 m: far_bounded (>= 40 m), never sky

data-007

What: synthetic (path-traced) procedural scenes. Structure: multi-view clips (views stored inside each clip).
Supervision tier: depth exact, cameras exact (supervision target).
Size: 2 clips, 48 frames, 0.1 GB; median 24 frames/clip; resolution 320x180; fps 24.0.
Modalities (share of clips): rgb 100%, depth 100%, normals 100%, instance 100%, semantic 100%, flow 100%, tracks 100%.
Tracks: ground-truth tracks with per-view visibility.
Flow: optical flow from the renderer.
Depth: transparent surfaces = first_surface (documented); mirrors unknown; volumetrics none; multi-layer: no (could be re-rendered).
Camera motion (index cam_*): fast 2; median image motion 3 deg/frame; shaky 2 (100 %).
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dynamic_tracks': 1.0, 'dense_motion': 1.0, 'covisibility': 1.0}.
test sample only (2 clips)
refractive materials marked in extra.ambiguity

data-008

What: real object-centric phone videos, multi-view-stereo depth. Structure: video clips.
Supervision tier: depth coarse, cameras estimated (conditioning input / guess for a synthetic-trained model; exact quantities only where the tier says exact).
Size: 49 clips, 5,299 frames, 5.4 GB; median 128 frames/clip; resolution 312x556, 340x453, 353x471; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, instance 100%, tracks 100%.
Tracks: static tracks from depth + cameras (scene is static).
Flow: none.
Depth: transparent surfaces = estimated (verified); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): moderate 6, fast 43; median image motion 3.1 deg/frame; shaky 39 (80 %).
Quality score: 1/5; metric scale: False; suggested weights: {'depth': 0.3, 'camera': 0.4, 'static_tracks': 0.3, 'covisibility': 0.3}.
lowest quality score: estimated cameras + fragmentary depth; heavily filtered to the best clips
arbitrary scale (scale-invariant losses only)
scene point clouds stored as scene assets

data-009

What: real RGB-D phone captures (sensor depth). Structure: video clips.
Supervision tier: depth sensor, cameras sensor (conditioning input / guess for a synthetic-trained model; exact quantities only where the tier says exact).
Size: 914 clips, 98,639 frames, 49.3 GB; median 128 frames/clip; resolution 480x640; fps 30.0.
Modalities (share of clips): rgb 100%, depth 100%, instance 100%, tracks 100%.
Tracks: static tracks (not on hand-held objects).
Flow: none.
Depth: transparent surfaces = sensor (documented); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): static 22, rotation_only 2, slow 26, moderate 438, fast 426; median image motion 2 deg/frame; shaky 232 (25 %).
Quality score: 2/5; metric scale: True; suggested weights: {'depth': 0.7, 'camera': 0.5, 'static_tracks': 0.4, 'covisibility': 0.4}.
estimated (AR) poses drift over long ranges: prefer windows <= 32 frames; drifting clips removed
sensor depth: holes and edge_uncertain bands

data-010

What: synthetic (rendered) dynamic people/animals. Structure: video clips; stereo pairs (left/right) linked by sync groups.
Supervision tier: depth exact, cameras exact (supervision target).
Size: 24 clips, 2,400 frames, 8.7 GB; median 128 frames/clip; resolution 1280x720; fps 30.0.
Modalities (share of clips): rgb 100%, depth 100%, instance 100%, flow 100%, tracks 100%.
Tracks: ground-truth trajectories (left camera); right-camera visibility derived by depth test.
Flow: forward/backward optical flow (1/16 px quantised).
Depth: transparent surfaces = first_surface (assumed); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): slow 24; median image motion 0.16 deg/frame; shaky 0 (0 %).
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dynamic_tracks': 1.0, 'dense_motion': 1.0, 'covisibility': 1.0}.
only one source archive converted so far (24 clips)

data-011

What: synthetic (game engine) street scenes, static. Structure: video clips.
Supervision tier: depth exact, cameras exact (supervision target).
Size: 120 clips, 12,000 frames, 10.6 GB; median 100 frames/clip; resolution 960x540; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, tracks 100%.
Tracks: static tracks from depth + cameras.
Flow: none.
Depth: transparent surfaces = first_surface (verified); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): moderate 5, fast 115; median image motion 3.5 deg/frame; shaky 7 (6 %).
Quality score: 4/5; metric scale: False; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'covisibility': 1.0}.
metric scale under review (possible large scale factor): scale-invariant losses only
sky = infinite depth (heuristic, horizon test)

data-012

What: synthetic (rendered) dynamic object scenes. Structure: short video clips (24 frames).
Supervision tier: depth exact, cameras exact (supervision target).
Size: 1,581 clips, 37,944 frames, 54.7 GB; median 24 frames/clip; resolution 512x512; fps 12.0.
Modalities (share of clips): rgb 100%, depth 100%, normals 100%, instance 100%, flow 100%, tracks 100%.
Tracks: static tracks + exact dynamic-object tracks (from stored object coordinates and per-frame object boxes).
Flow: forward/backward optical flow in the standard convention (corrected from the source's (dy, dx)).
Depth: transparent surfaces = first_surface (documented); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): static 73, rotation_only 12, slow 414, moderate 967, fast 115; median image motion 0.8 deg/frame; shaky 2 (0 %).
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dense_motion': 1.0, 'covisibility': 1.0}.
background dome at 40 m: far_bounded, never sky
instance semantics: object category ids in instances.json

data-013

What: synthetic (game engine) multi-camera dynamic scenes with humans and objects. Structure: NOT p4d shards: per-(sequence, camera) source units read by the repo's streaming loader for this code, which builds clips on the fly; humans/ holds per-person body + clothing vertex caches the loader range-reads.
Supervision tier: depth exact, cameras exact (supervision target).
Tracks: dense face-id warped tracks for environment, objects and the person (body instance 1, clothing instance 2); visibility by depth test (max(3 %, 3 cm)) + depth_residual.
Flow: none.
Depth: transparent surfaces = last_opaque (verified); mirrors unknown; volumetrics none; multi-layer: no.
Quality score: 4/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dynamic_tracks': 1.0, 'covisibility': 1.0}.
8 synchronized cameras per sequence (4 in the simulation subset)
far depth >= 655 m stored as far_bounded
loader outputs a right-handed world
body and clothing meshes are the source's; hair and shoes have no mesh (body-labelled pixels there are 1-2 cm in front of the body mesh)

data-014

What: real game footage with game depth buffer and estimated cameras. Structure: video clips.
Supervision tier: depth exact, cameras estimated (conditioning input / guess for a synthetic-trained model; exact quantities only where the tier says exact).
Size: 210 clips, 16,559 frames, 30.7 GB; median 72.5 frames/clip; resolution 1280x720; fps 24.0, 30.0.
Modalities (share of clips): rgb 100%, depth 100%, tracks 100%.
Tracks: static tracks (same-time only).
Flow: none.
Depth: transparent surfaces = mixed (documented); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): rotation_only 1, slow 18, moderate 92, fast 99; median image motion 1.9 deg/frame; shaky 174 (83 %).
Quality score: 3/5; metric scale: True; suggested weights: {'depth': 0.6, 'camera': 0.6, 'static_tracks': 0.5, 'covisibility': 0.6}.
broken calibrations and low-consistency clips filtered
far/near clipped depth -> hole

data-015

What: synthetic (path-traced) indoor scenes with transparent objects. Structure: single images.
Supervision tier: depth exact, cameras intrinsics_only (conditioning input / guess for a synthetic-trained model; exact quantities only where the tier says exact).
Size: 15,295 clips, 15,295 frames, 17.2 GB; median 1 frames/clip; resolution 1280x720; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%.
Tracks: none.
Flow: none.
Depth: transparent surfaces = first_surface (documented); mirrors unknown; volumetrics none; multi-layer: 8 layers.
Camera motion (index cam_*): single 15,295.
Quality score: 4/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 0.8, 'covisibility': 0.8}.
MULTI-LAYER DEPTH: 8 layers (depth_layers), one per media transition; layer 1 = standard depth
intrinsics derived from the generator config (camera weight 0.8)

data-016

What: synthetic (path-traced) indoor scenes. Structure: unordered views per trajectory (not smooth video).
Supervision tier: depth exact, cameras exact (supervision target).
Size: 769 clips, 74,587 frames, 102.2 GB; median 100 frames/clip; resolution 1024x768; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, normals 100%, instance 100%, semantic 100%, tracks 100%.
Tracks: static tracks from depth + cameras.
Flow: none.
Depth: transparent surfaces = first_surface (verified); mirrors surface; volumetrics none; multi-layer: no.
Transparent labels: nyu40 ids {'9': 'window', '19': 'mirror'} (no glass class; glass cups/bottles fall in 40 otherprop; 13 blinds / 28 showercurtain partly see-through).
Camera motion (index cam_*): unordered 769.
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'dense_motion': 1.0, 'covisibility': 1.0}.
tilt-shift cameras represented exactly (K may have skew)
windows/mirrors: first surface, labelled transparent
back-facing source normals flipped toward the camera

data-017

What: aerial imagery: real (photogrammetry mesh or LiDAR depth) + synthetic renders. Structure: single images.
Supervision tier: depth mixed, cameras intrinsics_only (conditioning input / guess for a synthetic-trained model; exact quantities only where the tier says exact).
Size: 16,214 clips, 16,214 frames, 10.5 GB; median 1 frames/clip; resolution 790x1198, 896x672, 952x532; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%.
Tracks: none.
Flow: none.
Depth: transparent surfaces = estimated (documented); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): single 16,214.
Quality score: 3/5; metric scale: True; suggested weights: {'depth': 0.8, 'camera': 0.6, 'covisibility': 0.6}.
no camera poses (identity extrinsics); per-image intrinsics
extra_meta.depth_kind = mesh | lidar | synthetic
real depth: edge_uncertain bands; LiDAR is sparse

data-018

What: synthetic (game engine) city, aerial + street views. Structure: unordered views.
Supervision tier: depth exact, cameras exact (supervision target).
Size: 142 clips, 17,402 frames, 16.6 GB; median 128 frames/clip; resolution 1000x1000, 1920x1080; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, tracks 100%.
Tracks: static tracks from depth + cameras.
Flow: none.
Depth: transparent surfaces = first_surface (verified); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): unordered 142.
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'covisibility': 1.0}.
street depth saturates at 655 m: far_bounded / sky
street views include upward-looking cameras

data-019

What: synthetic (game engine) outdoor environments. Structure: video clips (easy + hard trajectories).
Supervision tier: depth exact, cameras exact (supervision target).
Size: 1,020 clips, 124,219 frames, 109.3 GB; median 128 frames/clip; resolution 640x640; fps not recorded (frame-index time).
Modalities (share of clips): rgb 100%, depth 100%, tracks 100%.
Tracks: static tracks from depth + cameras.
Flow: none.
Depth: transparent surfaces = last_opaque (verified); mirrors unknown; volumetrics none; multi-layer: no.
Camera motion (index cam_*): rotation_only 1, moderate 1, fast 1,018; median image motion 4.8 deg/frame; shaky 1,015 (100 %).
Quality score: 5/5; metric scale: True; suggested weights: {'depth': 1.0, 'camera': 1.0, 'static_tracks': 1.0, 'covisibility': 1.0}; per-scene weights: {'ForestEnv': {'static_tracks': 0.8}, 'SeasonalForestSpring': {'static_tracks': 0.8}, 'SeasonalForestWinter': {'static_tracks': 0.8}, 'SeasonalForestAutumn': {'static_tracks': 0.8}, 'SeasonalForestWinterNight': {'static_tracks': 0.8}, 'GreatMarsh': {'static_tracks': 0.8}}.
foliage moves: static-track weight 0.8 for vegetation-heavy environments
sky ids per environment (some environments use two)

data-020

What: real phone / camera video of real-world scenes (mostly outdoor), photometric supervision; multi-view-stereo depth (third-party reconstruction, relative scale) where present. Structure: video clips (~4-6 fps, <= 128 frames; consistent segments of each scene).
Supervision tier: depth coarse, cameras estimated (conditioning input / guess for a synthetic-trained model; exact quantities only where the tier says exact).
Size: 6,276 clips, 659,313 frames, 340.8 GB; median 107 frames/clip; resolution 480x270, 960x540; fps 0.6674, 1.0013, 1.3031, 1.4302, 1.5008, 1.5804, 1.6218, 1.6222, 1.624, 1.7685, 1.8679, 1.8686, 1.8692, 1.8707, 1.8732, 1.8765, 1.9937, 1.9944, 1.996, 2.0061, 2.068, 2.0707, 2.1355, 2.1362, 2.1365, 2.1428, 2.1445, 2.1449, 2.1455, 2.2251, 2.2252, 2.3016, 2.3081, 2.3091, 2.3099, 2.3123, 2.3129, 2.313, 2.3131, 2.3132, 2.4015, 2.4907, 2.4908, 2.5, 2.5002, 2.5008, 2.5012, 2.5015, 2.5017, 2.502, 2.5025, 2.5035, 2.5038, 2.5054, 2.6085, 2.6109, 2.6122, 2.7162, 2.7168, 2.7173, 2.7178, 2.7213, 2.7239, 2.7252, 2.7279, 2.7287, 2.7299, 2.7304, 2.7308, 2.7312, 2.7317, 2.7321, 2.7324, 2.733, 2.7334, 2.7345, 2.7585, 2.8571, 2.8598, 2.8615, 2.8633, 2.9904, 2.9912, 2.992, 2.9925, 2.9933, 2.9954, 2.999, 2.9998, 3.0022, 3.0026, 3.0032, 3.0037, 3.0044, 3.0046, 3.0047, 3.0051, 3.0054, 3.0056, 3.0079, 3.0082, 3.1572, 3.1573, 3.1593, 3.1602, 3.1603, 3.1623, 3.1627, 3.1648, 3.1656, 3.3126, 3.319, 3.3191, 3.3198, 3.32, 3.3208, 3.3215, 3.3217, 3.3218, 3.3226, 3.3233, 3.3237, 3.325, 3.3255, 3.3268, 3.3301, 3.3309, 3.3317, 3.3325, 3.3328, 3.3331, 3.3333, 3.3336, 3.3339, 3.3341, 3.3344, 3.3347, 3.3349, 3.335, 3.3357, 3.3363, 3.3364, 3.3366, 3.3369, 3.337, 3.3371, 3.3372, 3.3376, 3.3378, 3.338, 3.3381, 3.3384, 3.3385, 3.3389, 3.339, 3.3399, 3.3402, 3.3405, 3.3411, 3.3412, 3.3427, 3.4857, 3.5265, 3.5278, 3.5281, 3.5282, 3.5283, 3.5288, 3.5294, 3.5296, 3.5301, 3.5302, 3.5306, 3.5307, 3.5308, 3.531, 3.5311, 3.5313, 3.5315, 3.5316, 3.532, 3.5324, 3.5326, 3.5327, 3.5329, 3.5335, 3.5336, 3.5337, 3.5338, 3.5339, 3.534, 3.5341, 3.5342, 3.5343, 3.5344, 3.5346, 3.5351, 3.5352, 3.5358, 3.5363, 3.5365, 3.5366, 3.5369, 3.5374, 3.658, 3.6642, 3.735, 3.7351, 3.7353, 3.7361, 3.7362, 3.7365, 3.7373, 3.7374, 3.7377, 3.7382, 3.7388, 3.739, 3.7393, 3.7394, 3.7407, 3.7409, 3.741, 3.7413, 3.7414, 3.7424, 3.7441, 3.7446, 3.7456, 3.7461, 3.7466, 3.7469, 3.7474, 3.7481, 3.7482, 3.7483, 3.7484, 3.7486, 3.7487, 3.749, 3.7494, 3.7495, 3.7496, 3.7497, 3.75, 3.7501, 3.7502, 3.7503, 3.7504, 3.7505, 3.7506, 3.7507, 3.7508, 3.7509, 3.751, 3.7511, 3.7512, 3.7513, 3.7514, 3.7515, 3.7516, 3.7518, 3.7519, 3.752, 3.7521, 3.7522, 3.7523, 3.7524, 3.7525, 3.7527, 3.753, 3.7531, 3.7532, 3.7533, 3.7534, 3.7535, 3.7536, 3.7538, 3.7539, 3.754, 3.7541, 3.7543, 3.7544, 3.7545, 3.7546, 3.7547, 3.7548, 3.7549, 3.755, 3.7551, 3.7553, 3.7554, 3.7555, 3.7556, 3.7559, 3.7561, 3.7562, 3.7563, 3.7564, 3.7567, 3.7568, 3.7569, 3.757, 3.7572, 3.7573, 3.7574, 3.7575, 3.7581, 3.7582, 3.7585, 3.7586, 3.7588, 3.7589, 3.759, 3.7593, 3.8304, 3.9589, 3.9964, 3.9966, 3.9973, 3.9974, 3.9978, 3.9982, 3.9983, 3.9986, 3.9987, 3.9991, 3.9995, 3.9996, 4.0, 4.0004, 4.0005, 4.0008, 4.0009, 4.0013, 4.0014, 4.0017, 4.0018, 4.0019, 4.0021, 4.0022, 4.0025, 4.0026, 4.0027, 4.003, 4.0031, 4.0033, 4.0034, 4.0036, 4.0037, 4.0038, 4.0039, 4.004, 4.0041, 4.0042, 4.0044, 4.0046, 4.0047, 4.0048, 4.0049, 4.005, 4.0051, 4.0052, 4.0053, 4.0054, 4.0056, 4.0059, 4.006, 4.0061, 4.0062, 4.0064, 4.0065, 4.0066, 4.0067, 4.0068, 4.0069, 4.007, 4.0071, 4.0072, 4.0073, 4.0075, 4.0078, 4.008, 4.0081, 4.0084, 4.0085, 4.0086, 4.0087, 4.0088, 4.0092, 4.0103, 4.1305, 4.2508, 4.2656, 4.2671, 4.2674, 4.2675, 4.2676, 4.2683, 4.2689, 4.269, 4.2698, 4.2702, 4.2704, 4.2708, 4.271, 4.2718, 4.2722, 4.2724, 4.2725, 4.2726, 4.2728, 4.2729, 4.2731, 4.2733, 4.2735, 4.2736, 4.2738, 4.2739, 4.274, 4.2741, 4.2745, 4.2747, 4.2748, 4.2753, 4.2754, 4.2762, 4.2763, 4.2766, 4.2769, 4.2772, 4.2774, 4.2781, 4.2783, 4.2797, 4.2799, 4.2804, 4.2815, 4.2816, 4.2827, 4.2828, 4.2836, 4.2837, 4.2838, 4.2839, 4.284, 4.2844, 4.2846, 4.2847, 4.2848, 4.2849, 4.2852, 4.2853, 4.2854, 4.2855, 4.2856, 4.2857, 4.2861, 4.2862, 4.2863, 4.2864, 4.2865, 4.2866, 4.2868, 4.2869, 4.287, 4.2871, 4.2873, 4.2875, 4.2877, 4.2879, 4.288, 4.2881, 4.2882, 4.2883, 4.2884, 4.2886, 4.2889, 4.289, 4.2891, 4.2892, 4.2896, 4.2897, 4.2898, 4.2899, 4.29, 4.2901, 4.2905, 4.2906, 4.2907, 4.2909, 4.2912, 4.2913, 4.2914, 4.2915, 4.2916, 4.2917, 4.2919, 4.292, 4.2921, 4.2922, 4.2923, 4.2924, 4.2925, 4.2926, 4.2928, 4.293, 4.2931, 4.2933, 4.2934, 4.2935, 4.2936, 4.2937, 4.2938, 4.2939, 4.294, 4.2943, 4.2944, 4.2945, 4.2946, 4.2949, 4.2952, 4.2953, 4.2954, 4.2955, 4.2956, 4.2957, 4.2959, 4.2961, 4.2962, 4.2964, 4.2965, 4.2966, 4.2969, 4.2971, 4.2975, 4.2981, 4.2983, 4.4778, 4.5076, 4.5117, 4.5324, 4.5409, 4.571, 4.6055, 4.6129, 4.613, 4.6131, 4.6132, 4.6134, 4.6136, 4.6141, 4.6143, 4.6144, 4.6145, 4.6147, 4.6148, 4.615, 4.6152, 4.6154, 4.6156, 4.6158, 4.616, 4.6162, 4.6163, 4.6164, 4.6165, 4.6166, 4.6167, 4.6168, 4.6169, 4.617, 4.6171, 4.6172, 4.6173, 4.6174, 4.6175, 4.6177, 4.6178, 4.6179, 4.618, 4.6181, 4.6182, 4.6183, 4.6184, 4.6185, 4.6186, 4.6187, 4.6188, 4.619, 4.6191, 4.6192, 4.6193, 4.6194, 4.6195, 4.6196, 4.6197, 4.6198, 4.6199, 4.62, 4.6201, 4.6202, 4.6203, 4.6204, 4.6205, 4.6206, 4.6208, 4.6209, 4.621, 4.6211, 4.6212, 4.6213, 4.6214, 4.6215, 4.6216, 4.6218, 4.6219, 4.622, 4.6221, 4.6223, 4.6224, 4.6225, 4.6226, 4.6227, 4.6228, 4.6229, 4.623, 4.6231, 4.6233, 4.6234, 4.6235, 4.6236, 4.6237, 4.6238, 4.624, 4.6241, 4.6242, 4.6245, 4.6246, 4.6247, 4.6248, 4.6249, 4.6251, 4.6252, 4.6253, 4.6254, 4.6255, 4.6256, 4.6257, 4.6258, 4.6259, 4.626, 4.6261, 4.6262, 4.6263, 4.6264, 4.6265, 4.6268, 4.6269, 4.627, 4.6271, 4.6273, 4.6274, 4.6277, 4.6278, 4.6279, 4.6281, 4.6282, 4.6284, 4.6285, 4.6286, 4.6295, 4.6304, 4.6556, 4.6777, 4.8002, 4.8117, 4.8122, 4.8968, 4.9183, 4.9283, 4.9336, 4.9579, 4.9766, 4.9781, 4.9784, 4.9786, 4.9788, 4.9789, 4.9799, 4.9807, 4.9809, 4.9811, 4.9813, 4.9816, 4.9827, 4.9829, 4.9832, 4.9835, 4.9841, 4.9843, 4.9849, 4.985, 4.9851, 4.9852, 4.9854, 4.9855, 4.9858, 4.9859, 4.986, 4.9862, 4.9867, 4.9873, 4.9874, 4.9875, 4.9876, 4.988, 4.9881, 4.9886, 4.9887, 4.9892, 4.9896, 4.9899, 4.9901, 4.9907, 4.991, 4.9915, 4.9916, 4.9926, 4.9936, 4.9937, 4.9941, 4.9942, 4.9947, 4.9953, 4.9954, 4.9955, 4.9956, 4.9959, 4.9968, 4.9969, 4.997, 4.9975, 4.9976, 4.9977, 4.998, 4.9983, 4.9985, 4.9992, 4.9993, 5.0, 5.0007, 5.0008, 5.0015, 5.0016, 5.0017, 5.0022, 5.0023, 5.0024, 5.0025, 5.0028, 5.003, 5.0031, 5.0032, 5.0033, 5.0037, 5.0038, 5.0039, 5.004, 5.0041, 5.0042, 5.0045, 5.0046, 5.0047, 5.0048, 5.0049, 5.005, 5.0052, 5.0053, 5.0054, 5.0055, 5.0056, 5.0057, 5.0059, 5.006, 5.0061, 5.0062, 5.0063, 5.0064, 5.0065, 5.0066, 5.0067, 5.0068, 5.0069, 5.007, 5.0071, 5.0072, 5.0073, 5.0074, 5.0075, 5.0076, 5.0077, 5.0078, 5.008, 5.0081, 5.0082, 5.0083, 5.0085, 5.0086, 5.0087, 5.0088, 5.0089, 5.009, 5.0091, 5.0093, 5.0094, 5.0095, 5.0096, 5.0097, 5.0098, 5.0099, 5.01, 5.0101, 5.0102, 5.0103, 5.0104, 5.0105, 5.0106, 5.0107, 5.0108, 5.0109, 5.011, 5.0111, 5.0112, 5.0113, 5.0114, 5.0115, 5.0116, 5.0117, 5.0118, 5.0119, 5.012, 5.0121, 5.0122, 5.0123, 5.0124, 5.0125, 5.0126, 5.0128, 5.0129, 5.013, 5.0131, 5.0132, 5.0133, 5.0136, 5.0137, 5.0138, 5.0139, 5.014, 5.0141, 5.0142, 5.0143, 5.0144, 5.0145, 5.0146, 5.0147, 5.0148, 5.0149, 5.0152, 5.0153, 5.0154, 5.0155, 5.0156, 5.0188, 5.1028, 5.167, 5.2444, 5.2567, 5.258, 5.2701, 5.4115, 5.4517, 5.4533, 5.4539, 5.4542, 5.4545, 5.4552, 5.4555, 5.4558, 5.4561, 5.457, 5.4571, 5.4573, 5.4576, 5.4577, 5.458, 5.4582, 5.4586, 5.4589, 5.4594, 5.4595, 5.4597, 5.4599, 5.46, 5.4601, 5.4603, 5.4604, 5.4606, 5.4613, 5.4617, 5.462, 5.4629, 5.4631, 5.4633, 5.4634, 5.4635, 5.4636, 5.4637, 5.464, 5.4645, 5.4646, 5.4648, 5.465, 5.4652, 5.4654, 5.4656, 5.4659, 5.466, 5.4664, 5.4667, 5.4673, 5.4677, 5.4678, 5.4679, 5.4681, 5.4682, 5.4683, 5.4684, 5.4686, 5.4688, 5.4691, 5.4694, 5.4698, 5.4701, 5.4702, 5.4703, 5.4706, 5.4786, 5.6099, 5.6707, 5.9737, 5.9746, 5.9809, 5.9825, 5.984, 5.9845, 5.9867, 5.9872, 5.9928, 5.9939, 5.9944, 5.9997, 6.0011, 6.0031, 6.0069, 6.0077, 6.0082, 6.0107, 6.0123, 6.1364, 6.1447, 7.2163.
Modalities (share of clips): rgb 100%, tracks 100%.
Tracks: sparse structure-from-motion points of the source reconstruction as static tracks; valid = visible only where the point was observed; 2-D observations in extra.sfm_observations.npz.
Flow: none.
Depth: transparent surfaces = estimated (verified); mirrors unknown; volumetrics none; multi-layer: no.
Quality score: 3/5; metric scale: False; suggested weights: {'depth': 0.15, 'camera': 0.7, 'static_tracks': 0.5, 'covisibility': 0.4, 'photometric': 1.0}.
native radial-tangential camera (camera_model 'opencv'); arbitrary scale
depth (where modalities.depth.present): multi-view-stereo camera z in the camera's units (not metres), resampled to the native distorted frame and stored at HALF resolution (modalities.depth.stride = 2; the reader upsamples to the RGB size); noisy and ~40-50 % valid: weight every pixel by depth_confidence (low near edges); sky (heuristic, large components only) supervised as infinite with the per-pixel extra.sky_weight
use for novel-view / photometric (Gaussian-splatting) losses + camera supervision; held-out views every 8th scene frame (extra_meta.nvs_split)
RGB is lossless WebP (pixels identical to the source PNG)
transient movers (people, cars) not masked; scene-level reflection / transparency labels in extra_meta.labels

Downloads last month: 635

Total file size:

2.66 TB