Datasets:
docs: dataset card with nuScenes attribution
Browse files
README.md
CHANGED
|
@@ -37,8 +37,18 @@ thumbs.tar -> repo root <- fetch this one
|
|
| 37 |
|---|---|---|
|
| 38 |
| `artifacts/vectors.npy` | 21 MB | 6858 × 768 float32, L2-normalised SigLIP2 image embeddings |
|
| 39 |
| `artifacts/caption_vectors.npy` | 10.5 MB | 6858 × 384 float32, MiniLM embeddings of generated captions |
|
| 40 |
-
| `thumbs.tar` |
|
| 41 |
-
| `*.jpg` |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
|
| 43 |
**The thumbnails are redacted.** Faces and licence plates are mosaicked at source resolution before
|
| 44 |
the downscale — 5,511 regions across the corpus, 2,616 frames (38.1%) carrying at least one.
|
|
@@ -64,9 +74,11 @@ unauthenticated caller for. It returned 429 at roughly the 490th file on a CI ru
|
|
| 64 |
minutes in, leaving a half-populated directory. One request cannot half-arrive.
|
| 65 |
|
| 66 |
The tarball is uncompressed deliberately: measured over 200 of these thumbnails, gzip -6 saves 2.1%,
|
| 67 |
-
because JPEG is already DCT-compressed. It is larger than the sum of its contents (
|
| 68 |
-
because tar pads every member to a 512-byte boundary
|
| 69 |
-
|
|
|
|
|
|
|
| 70 |
|
| 71 |
Row order in both matrices matches `frames.jsonl` in the application repo exactly, and the
|
| 72 |
application re-checks that on load — a misaligned row would mean confident answers about the wrong
|
|
|
|
| 37 |
|---|---|---|
|
| 38 |
| `artifacts/vectors.npy` | 21 MB | 6858 × 768 float32, L2-normalised SigLIP2 image embeddings |
|
| 39 |
| `artifacts/caption_vectors.npy` | 10.5 MB | 6858 × 384 float32, MiniLM embeddings of generated captions |
|
| 40 |
+
| `thumbs.tar` | 267 MB | all 6,858 thumbnails, flat and uncompressed, faces/plates redacted |
|
| 41 |
+
| `*.jpg` | 261 MB | the same thumbnails loose, 640×360, downscaled from 1600×900 |
|
| 42 |
+
|
| 43 |
+
**640×360 as of 2026-08-06, regenerated from the nuScenes source.** Thumbnail width is one constant,
|
| 44 |
+
`scripts/anonymise.py: THUMB_WIDTH`, raised off 320 on 2026-08-05; this published set predated that
|
| 45 |
+
change by a day and has now caught up with it. 320 was inherited from nuScenes and discards most of
|
| 46 |
+
the frame on modern footage, which is a 2× rather than 5× downscale.
|
| 47 |
+
|
| 48 |
+
The rewrite re-read the source images and re-applied redaction at source resolution, so **the privacy
|
| 49 |
+
figures below are reproduced exactly rather than carried over** — same 5,511 regions, same 2,616
|
| 50 |
+
frames, same 239/348/4,924 split, `privacy_unverified` still 0. Redaction geometry does not depend on
|
| 51 |
+
the output width, which is the point of mosaicking before the downscale.
|
| 52 |
|
| 53 |
**The thumbnails are redacted.** Faces and licence plates are mosaicked at source resolution before
|
| 54 |
the downscale — 5,511 regions across the corpus, 2,616 frames (38.1%) carrying at least one.
|
|
|
|
| 74 |
minutes in, leaving a half-populated directory. One request cannot half-arrive.
|
| 75 |
|
| 76 |
The tarball is uncompressed deliberately: measured over 200 of these thumbnails, gzip -6 saves 2.1%,
|
| 77 |
+
because JPEG is already DCT-compressed. It is larger than the sum of its contents (267 MB against
|
| 78 |
+
261) because tar pads every member to a 512-byte boundary. **That overhead used to be the notable
|
| 79 |
+
part of this paragraph and no longer is** — at 320×180 the members averaged ~10 KB and padding cost
|
| 80 |
+
17%; at 640×360 they average 37 KB and it costs 2%. Padding is a fixed ~0.75 KB per member, so it
|
| 81 |
+
matters in inverse proportion to file size.
|
| 82 |
|
| 83 |
Row order in both matrices matches `frames.jsonl` in the application repo exactly, and the
|
| 84 |
application re-checks that on load — a misaligned row would mean confident answers about the wrong
|