Silicon23/itw_pipeline_visualization

ITW Pipeline Visualization — assets

Rendered assets for the visualization site of an in-the-wild open-vocabulary 3D video object tracking project. Each pipeline takes an RGB video plus a text category label and produces a per-frame 3D bounding-box trajectory; the site shows what each one made of the same clip.

This repository holds only the heavy media the site streams at runtime: per-frame depth and colour for the browser-side point cloud, overlay renders, and gallery previews. The site itself is static and hosted separately; it fetches these files cross-origin.

These are derived visualization artifacts, not training data. Every frame originates from a source dataset — see Datasets and attribution.

Datasets and attribution

Assets are grouped by source dataset. Each dataset's assets carry that dataset's own licence and attribution. More datasets will be added; this table is the authoritative list of what is currently present.

id	source dataset	licence	attribution
sav	SA-V (Segment Anything Video, Meta AI)	CC BY 4.0	Ravi et al., SAM 2: Segment Anything in Images and Videos, 2024

Derived overlays and depth encodings are redistributed under the licence of the dataset they derive from. The repository-level licence tag reflects the sources currently included and will be revisited as datasets with different terms are added — always read the table above rather than the tag alone.

Layout

assets/<dataset_id>/<video_id>/
├── preview.mp4 · poster.jpg · strip.jpg        gallery preview
└── <pipeline_id>/
    ├── manifest.json          frames, intrinsics, metric camera-to-world poses
    ├── boxes.json             per-object per-frame 3D boxes (see conventions)
    ├── visualizations.json    what overlays exist, and at what cadence
    ├── depth/000000.webp …    uint16 metric depth, packed into R and G
    ├── rgb/000000.jpg …       colour for the point cloud
    └── overlays/<name>/
        ├── video.mp4          for playback
        └── frames/NNNNNN.jpg  stills at that overlay's own cadence

Pipelines currently present: wilddet3d (prompt-and-detect). Overlays currently present: input (source video with ground-truth masks), box_prompt, point_prompt.

Conventions

These hold across every dataset and pipeline in this repository.

Depth. WebP with the 16-bit value split across the red and green channels, because no single-channel 16-bit format decodes reliably across browsers:

z_metres = (R * 256 + G) / 65535 * depth_max_m      # z == 0 means invalid

depth_max_m is per clip, in manifest.json. Depth is metric metres, camera-Z, OpenCV axes (x right, y down, z forward).

Boxes. Camera frame — the detector applies no extrinsic. Multiply by the per-frame c2w in manifest.json for world frame. Each box carries center, dims_xyz (extents along the box's own axes), quat_wxyz, and corners.

Corner order is canonical and identical across pipelines:

0 (-,-,-)  1 (+,-,-)  2 (+,+,-)  3 (-,+,-)
4 (-,-,+)  5 (+,-,+)  6 (+,+,+)  7 (-,+,+)

Edges: bottom 0-1-2-3-0, top 4-5-6-7-4, verticals 0-4 1-5 2-6 3-7. Corners are validated elementwise against center + (signs · dims/2) @ Rᵀ before export, because IoU is order-invariant and cannot catch an ordering mistake.

Frame indices are always native source-video frames, so overlays at different cadences share one timeline. Every overlay video is encoded at the source frame rate with its content held between updates, so all overlays for a clip have identical duration and stay in sync under a single playback rate.

Annotation cadence is preserved, not interpolated. Where a source dataset annotates below video rate — SA-V labels every 4th frame at 24 fps, giving 6 Hz masks — the overlay holds the previous annotation rather than inventing intermediate ground truth, and visualizations.json records the true cadence.