LTtestMay10 / README.md
luuuulinnnn's picture
upload README.md
21540cc verified

LTtestMay10 — per-clip stride=1 30 fps test set

218 self-contained clips for pose / depth evaluation. Each clip is 41 frames at 30 fps stride=1 (1.37 s real time, 2.56 s when played at 16 fps), undistorted to a pinhole 832×480 model. All 6 surround cameras are included; DC depth is provided for the front camera only (matching the original LongtailTest).

Sourced from 126 KEEP UUIDs of NVIDIA's PhysicalAI-Autonomous-Vehicles dataset across chunks 234–237. The pipeline (download raw → ftheta undistort → DepthCrafter) is identical to the original LongtailTest, but applied at the native 30 fps cadence (no temporal downsampling).

Layout

chunk_234.tar          # all kept clips for chunk 234 (~1.5–2 GB each)
chunk_235.tar
chunk_236.tar
chunk_237.tar
manifest_clips.jsonl   # 218 entries  {chunk, uuid, clip_id, window_start, displacement_m}
kept_clips.txt         # 218 lines    chunk\tuuid\tclip_id
dropped_clips.jsonl    # 34 entries   (displacement < 2 m within 1.37 s)
README.md

After extracting all four chunk tars:

chunk_NNN/<uuid>/clip_NNNNNN/
├── front.mp4         # 41 frames, 832×480, h264 CRF 18, 30 fps
├── cross_left.mp4    # same spec
├── cross_right.mp4
├── rear_left.mp4
├── rear_right.mp4
├── rear_tele.mp4
├── front_depth.pt    # {'depth_sequence': (41, 1, 512, 832) fp16,
│                       'source_indices': [start..start+40]}
├── pose.pt           # {'T_anchor_front': (11, 4, 4) world_from_cam in OpenCV,
│                          first anchor = identity}  (FRONT camera),
│                       'T_anchor_all':   (6, 11, 4, 4) — same convention, one per view,
│                          ordered by sensor_order; T_anchor_all[0] == T_anchor_front,
│                       'sensor_order':   list of 6 camera names}
└── meta.pt           # K (front, 3,3), K_all (6, 3, 3) target pinhole intrinsics,
                        E_rig_front (4,4), E_all (6, 4, 4) sensor extrinsics,
                        sensor_order [6 names], view_files [6 mp4 names],
                        frame_indices, timestamps_us, anchor_clip_idx,
                        anchor_src_idx, anchor_t_us, src_fps=30, stride=1,
                        window_start, chunk, uuid, clip_id, anchor_displacement_m

Per-UUID windows

Two 41-frame windows per UUID at 30 fps source frames:

clip_id window_start end (exclusive) real time start
clip_000000 0 41 0.00 s
clip_000001 300 341 10.00 s

The two windows are 10 s apart in real time, providing diverse trajectories per UUID.

Pose anchors

T_anchor_front has 11 anchors at clip-frame indices 0, 4, 8, …, 40 (i.e. every 4th clip frame). At stride=1 from 30 fps, this equals real-time spacing ≈ 0.133 s, total span 1.37 s. Anchors are world_from_cam in OpenCV camera frame, with the first anchor forced to identity:

T_anchor_front[i] = inv(T_world_front[0]) @ T_world_front[i]
T_world_front(t)  = T_world_rig(t) @ E_rig_front

T_world_rig(t) is interpolated from egomotion.offline.parquet via SLERP (rotation) + linear (translation) at the 11 anchor camera timestamps. E_rig_front comes from sensor_extrinsics.offline.parquet.

meta.K is the target pinhole intrinsics (constant across UUIDs):

fx = 400.0, fy = 411.0, cx = 415.0, cy = 338.0   # output 832×480

Motion filter

Clips with ‖T_anchor_front[10][:3,3] − T_anchor_front[0][:3,3]‖ < 2.0 m are excluded from the tarballs (not present at all). The dropped 34 clip ids are listed in dropped_clips.jsonl for transparency. 124 of the 126 UUIDs have at least one clip that passes the filter; 2 UUIDs (in chunk 236) have both clips dropped (parked / heavy traffic).

Counts

chunk UUIDs clips total kept (≥ 2 m) dropped
234 30 60 (see manifest) (see dropped_clips)
235 35 70 (see manifest)
236 34 68 (see manifest)
237 27 54 (see manifest)
total 126 252 218 34

Loading example

from huggingface_hub import snapshot_download
import torch, av

snapshot_download("luuuulinnnn/LTtestMay10", repo_type="dataset",
                  local_dir="LTtestMay10", allow_patterns=["chunk_234.tar","kept_clips.txt"])
# extract chunk tars locally, then:

clip_dir = "LTtestMay10/chunk_234/<uuid>/clip_000000"
pose  = torch.load(f"{clip_dir}/pose.pt")["T_anchor_front"]      # (11, 4, 4)
depth = torch.load(f"{clip_dir}/front_depth.pt")["depth_sequence"]   # (41, 1, 512, 832) fp16
meta  = torch.load(f"{clip_dir}/meta.pt")                       # K, E_rig_front, ...
container = av.open(f"{clip_dir}/front.mp4")

Provenance

  • Raw video: nvidia/PhysicalAI-Autonomous-Vehicles chunks 0234–0237 (camera_front_wide_120fov, 30 fps, 1920×1080 ftheta)
  • KEEP UUIDs: filter on Qwen2.5-VL-7B captions (sunny / bright_day / clear); 126 UUIDs across 4 chunks
  • ftheta undistort: own reproduction of the original ray_wan pipeline (math verified pixel-equivalent within mpeg4 noise floor)
  • DepthCrafter: tencent/DepthCrafter, max_res=1024, 5 inference steps, window_size=110, overlap=25, run on 30 fps undistorted videos
  • Pose GT: SLERP+linear interpolation of egomotion.offline.parquet at per-clip anchor timestamps; mean translation diff vs original LongtailTest GT is ~1.4 m on ~80 m trajectories (≈ 1.7 %, within egomotion sub-frame timestamp accuracy)