P-057-RisingStar/scenesearch-assets

SceneSearch corpus assets

The two things SceneSearch needs at runtime that are too large to keep in git: the search index and the frame thumbnails, both covering 6,858 CAM_FRONT keyframes from 170 nuScenes v1.0-trainval scenes.

This is derived data, not a source. The authoritative record of a frame is its coordinate — video_id plus timestamp_s in the application's frames.jsonl, which is committed — and everything here is regenerable from nuScenes by re-running the project's ingest. Nothing is derivable only from this repo.

Contents

vectors.npy, caption_vectors.npy   ->  artifacts/
thumbs.tar                         ->  repo root      <- fetch this one
<frame_id>_CAM_FRONT.jpg           ->  repo root, flat

artifacts/vectors.npy	21 MB	6858 × 768 float32, L2-normalised SigLIP2 image embeddings
artifacts/caption_vectors.npy	10.5 MB	6858 × 384 float32, MiniLM embeddings of generated captions
thumbs.tar	274 MB	all 6,858 thumbnails, flat and uncompressed, faces/plates redacted
*.jpg	261 MB	the same thumbnails loose, 640×360, downscaled from 1600×900

640×360 as of 2026-08-06, regenerated from the nuScenes source. Thumbnail width is one constant, scripts/anonymise.py: THUMB_WIDTH, raised off 320 on 2026-08-05; this published set predated that change by a day and has now caught up with it. 320 was inherited from nuScenes and discards most of the frame on modern footage, which is a 2× rather than 5× downscale.

The rewrite re-read the source images and re-applied redaction at source resolution, so the privacy figures below are reproduced exactly rather than carried over — same 5,511 regions, same 2,616 frames, same 239/348/4,924 split, privacy_unverified still 0. Redaction geometry does not depend on the output width, which is the point of mosaicking before the downscale.

The thumbnails are redacted. Faces and licence plates are mosaicked at source resolution before the downscale — 5,511 regions across the corpus, 2,616 frames (38.1%) carrying at least one.

Regions come from two sources unioned together: dedicated face and licence-plate detectors (YOLO11n, 0.35 confidence floor) contributing 239 faces and 348 plates, and the top 28% of every person box contributing 4,924 head bands. Vehicle boxes produce no regions. The head band is not a fallback for a weak detector — measured on crowd footage the face detector is sound, and 68% of frames containing a person still yield no detectable face at all, so the union is what coverage actually rests on. Here it is 89% of the redacted area, because nuScenes ships pre-blurred and the dedicated detectors fire on only 7.5% of frames.

Coverage is therefore the union of two recalls, and this is a mitigation rather than a guarantee that no face is visible; nuScenes' own blurring pass sits underneath it. Every frame in this dataset is covered by a provenance record naming the exact weights that redacted it — privacy_detections.jsonl in the source repo's data/artifacts/, which scripts/verify_privacy_assets.py checks in both directions before anything is uploaded. scripts/anonymise.py has the geometry and the full list of limits.

Fetch thumbs.tar, not the loose files. They hold identical bytes, and the loose copies exist so the dataset is browsable here — but pulling 6,858 files one at a time is what the Hub rate-limits an unauthenticated caller for. It returned 429 at roughly the 490th file on a CI runner, four and a half minutes in, leaving a half-populated directory. One request cannot half-arrive.

The tarball is uncompressed deliberately: measured over 200 of these thumbnails, gzip -6 saves 2.1%, because JPEG is already DCT-compressed. It is larger than the sum of its contents — 274 MB against 261, an overhead of 12.3 MB or 4.7% — because every member carries header blocks and is padded to a 512-byte boundary. Measured over this archive that is 1,790 bytes per member: Python's tarfile writes PAX by default, so each file gets an extended-header pair on top of the ustar header, plus alignment padding.

That cost is fixed per member, so it matters in inverse proportion to file size — which is why this paragraph used to argue harder. The same ~12 MB was 17% of the 320×180 set (members averaged ~10 KB) and is 4.7% of this one (members average 37 KB). Nothing about the packing changed; the files got bigger.

Row order in both matrices matches frames.jsonl in the application repo exactly, and the application re-checks that on load — a misaligned row would mean confident answers about the wrong frames. Thumbnails are named <frame_id>_CAM_FRONT.jpg, where frame_id is the nuScenes sample_data token.

The thumbnails sit at the repo root rather than under a matching thumbs/ prefix only because moving 6,858 published files server-side would cost a full re-upload for cosmetics.

The 170 scenes were not chosen at random. They were selected so that rain and night appear independently rather than together — in the 10-scene subset this project started from, every rainy scene was also a night scene, which makes the two conditions impossible to tell apart when measuring. See the project's eval/results/report.md for the resulting measurements.

Usage

From a SceneSearch checkout, both are fetched by one script — and make run and make test call it already, so a fresh clone needs no extra step:

python scripts/fetch_assets.py              # index -> data/artifacts/, thumbnails -> data/thumbs/
python scripts/fetch_assets.py --what index # just what the service cannot start without

Directly:

import tarfile
from huggingface_hub import hf_hub_download

REPO = "P-057-RisingStar/scenesearch-assets"
hf_hub_download(REPO, "artifacts/vectors.npy", repo_type="dataset", local_dir="data/artifacts")

tar = hf_hub_download(REPO, "thumbs.tar", repo_type="dataset", local_dir="data")
with tarfile.open(tar, "r:") as t:
    t.extractall("data/thumbs", filter="data")   # filter= requires Python 3.11.4+

SceneSearch pulls both at image build time, never per request, so the running service reads from its own disk and does not depend on this repo being reachable. The index pull is fatal and the thumbnail pull is not: without the index the service cannot answer at all, while without thumbnails it answers correctly and shows broken images.

Licence and attribution

Derived from nuScenes by Motional, released under CC BY-NC-SA 4.0. These assets are a derivative work and carry the same licence: attribution, non-commercial, share-alike.

nuScenes: A multimodal dataset for autonomous driving. Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, Oscar Beijbom. CVPR 2020.

Non-commercial use only, per the upstream licence. If you need these images commercially, go to nuScenes directly — this repo cannot grant terms it does not hold.

Provenance

Produced by scripts/publish_assets.py in the SceneSearch repository. Regenerating them from a nuScenes install and re-running that script reproduces this repo exactly.