azur0licht/kiwi_example_traj

KIWI Example Trajectories

Processed 6-DoF trajectories from bimanual recordings: left and right hand-mounted camera poses and a head-mounted marker pose. Tracks within an episode share a metric coordinate frame and clock.

Trajectories for scene_001 / episode_001 and scene_002 / episode_001 are stored as Parquet in the example split. Camera recordings with audio removed are available under raw/ for scene_001 / episode_001 and scene_002 / episode_001.

Contents

Body	Median sampling rate (Hz)
left_front_camera	119.88
right_front_camera	119.88
head_marker	15.00

The scene_002 head positions include local interpolation. The method, intervals, and retained components are listed in metadata.json.

The hand trajectories describe the front camera optical frames. The head trajectory describes the head-mounted marker frame. Each row contains one pose of one body; samples are ordered by timestamp within each scene and episode. Each body has its own sampling grid, so use timestamps to align the tracks.

Episode numbering starts within each scene. Identify a recording by (scene_id, episode_id) and a separate time series by (scene_id, episode_id, body).

Format

Field	Type	Description
scene_id	string	Scene identifier (scene_001 or scene_002)
episode_id	string	Episode identifier within its scene (episode_001)
body	string	One of the three bodies listed above
sample_index	int64	Zero-based sample index within that scene, episode, and body
timestamp_s	float64	Time in seconds on the shared episode clock
position_m	float64[3]	World-frame position, ordered x, y, z, in metres
quaternion_xyzw	float64[4]	Hamilton quaternion, ordered x, y, z, w; rotates body coordinates into world coordinates

Each scene has its own metric, gravity-aligned world frame with +Z up and a scene-relative origin. Each (scene_id, episode_id) has its own clock; coordinates and timestamps from different scenes are independent. Camera axes follow the OpenCV convention: x right, y down, z forward. Timestamps are relative to the episode clock and may begin slightly before zero.

Load

Load with Hugging Face Datasets:

from datasets import load_dataset

poses = load_dataset("azur0licht/kiwi_example_traj", split="example")
scene = poses.filter(lambda row: row["scene_id"] == "scene_002")
head = scene.filter(lambda row: row["body"] == "head_marker")
head_arrays = head.with_format("numpy", dtype="float64")[:]

metadata.json provides frame definitions, sampling rates, and timestamp ranges. SHA256SUMS lists file checksums.

Camera recordings

Audio has been removed from all camera recordings. Original video encoding and timestamps are retained; hand-camera INSV files retain embedded IMU telemetry.

Scene / episode	Left camera	Right camera	Head camera
scene_001 / episode_001	left.insv	right.insv	head.mp4
scene_002 / episode_001	left.insv	right.insv	head.mp4

Each hand-camera INSV contains both fisheye video streams. Each head-camera MP4 contains its video stream. Both scenes have trajectory examples and their own Gaussian reconstruction.

The recordings are untrimmed and retain their native camera timelines. Video timestamps are not automatically on the shared trajectory clock. The head_marker trajectory describes the marker frame, rather than the head camera's optical frame.

metadata.json groups recordings by scene and episode and identifies their video streams. For each episode, it also provides hand-video timestamp offsets: timestamp_s = native_video_pts_s + trajectory_time_offset_s. These offsets apply to video frames. A null offset means no mapping is provided for that stream.

Download an individual recording with Hugging Face Hub:

from huggingface_hub import hf_hub_download

video_path = hf_hub_download(
    repo_id="azur0licht/kiwi_example_traj",
    repo_type="dataset",
    filename="raw/scene_001/episode_001/left.insv",
)

The recordings are separate assets; load_dataset(..., split="example") loads the trajectory table. File sizes and SHA-256 checksums are provided in metadata.json and SHA256SUMS.

Reconstructed scenes

Each scene has a separate Gaussian reconstruction, stored as a compressed SPZ version-4 asset with degree-3 spherical harmonics.

Scene	Gaussian reconstruction
scene_001	scene.spz
scene_002	scene.spz

Use an SPZ version-4-compatible reader. Both assets use metres and RUB coordinates (Y up). Trajectories use a Z-up world frame. The reconstructions list in metadata.json provides each asset's T_world_asset transform into its corresponding scene's trajectory world frame. The two scenes have independent origins.

For a viewer retaining the decoded SPZ's RUB axes, transform trajectory poses with T_asset_body = inverse(T_world_asset) @ T_world_body. The transform does not rescale the scene. Account for any additional coordinate conversion performed by your viewer.

The scenes are separate downloadable assets. The example split continues to contain the trajectory table.

Browser video previews

Untrimmed H.264/yuv420p MP4 previews are available under previews/<scene>/<episode>/. Left/right front and rear files are separate, unstitched fisheye lens streams; they are not rectified or panoramic renders. Previews are resized and frame-rate limited for responsive browser playback. Dimensions, output fps and compression settings are recorded in each receipt. They contain no audio or IMU. Original recordings remain unchanged; downsampled previews retain the episode clock mapping and complete recording duration, within one output frame.

Each preview has a provenance receipt containing the source revision/checksum, stream index, processing command, frame count, output checksum and preview encoding settings. episodes[].videos links previews and originals. For mapped streams, video_time = video_start_s + (timestamp_s - episode_start_s) * playback_rate. Head previews have timing: null: their synchronization offset is not supplied, so viewers must use independent playback rather than assume a zero offset.

These previews were generated from source revision bf118eec4b1977661940e28e8f123f8132d2f038, before the subsequent audio-removal update. They already contain no audio. Source checksums and revisions in the receipts identify the exact historical inputs; they are not checksums of the current raw files. The PR retains the current raw assets and their updated checksums.