Oratis's picture
完成归档核验并补齐逐文件哈希和分片索引
a454e7f verified
|
Raw History Blame Contribute Delete
4.98 kB
---
pretty_name: Diogenes CARLA Town10HD Multicamera Archive v01
license: other
license_name: original-capture-archive-terms
license_link: https://huggingface.co/datasets/DiogenesLab/Diogenes_CARLA_Multicamera_Archive_v01/blob/main/NOTICE.md
tags:
- video
- camera-pose
- archival
- non-trainable
size_categories:
- n<1K
---
# Diogenes CARLA Town10HD Multicamera Archive v01
The Diogenes project has closed. This public archive preserves original capture artifacts and historical limitations for inspection and reproducibility. **This archive manifest is fully transferred and verified.** `ARCHIVE_INDEX.json` lists the exact tar files and fixed verification revisions. This update's verified manifest covers **232 objects / 5,440,205,827 public bytes**, with **4 explicitly marked path-redacted copies**. All listed public members have passed source/archive SHA-256, pinned Hugging Face LFS size/hash, and anonymous download checks.
## Scope and historical status
The overall CARLA collection comprises six 60-second Town10HD_Opt four-camera groups from 2026-08-10 and ten 60-second earlier autopilot demo groups from 2026-08-07. Group `2026_08_10_carlatown_05` was already transferred as the verified pilot; it is intentionally absent from this update's source manifest. The remaining five later groups and ten early demos retain their original folder names. Shared historical staging documentation is retained as provenance and may reference other scenes that are not part of this repository.
The later batch originally reported 125 PASS / 0 FAIL / 9 SKIP, but subsequent delivery-contract review found unmet criteria; this archive does not repeat the original report as current delivery approval. Its format is 1920×1080, 30 fps, four 1,800-frame streams per 60-second sample. All later samples are 16:9, do not meet the requested aspect/subject quotas, and contain **synthetic ambient noise rather than CARLA-recorded sound**. Small occlusions and framing defects remain.
The early autopilot demo is explicitly a route-validation batch, not a compliant delivery batch. Its preserved example reports contain 17–20 failures: vehicle-motion derivatives exceed smooth-camera limits, starting viewpoints are too far apart, and input/audio checks fail. CARLA-native axes are left-handed x-forward / y-right / z-up; the record keeps quaternion xyzw. No geometry has been reprojected or corrected for this archival publication.
CARLA-specific asset attribution is retained. Upstream 0.9.15 identifies CC-BY without a version, so this card does not invent one; CARLA code uses MIT. This is not a grant of rights in unrelated Unreal or third-party components.
## Reading the archive
Files are grouped into ordinary uncompressed TAR shards under `archive/`. Each tar preserves the listed relative paths and ends with its own `MANIFEST.jsonl` containing exact member sizes and SHA-256. The repository-level `MANIFEST.jsonl` now contains the verified SHA-256 of every public member. `ARCHIVE_INDEX.json` records each tar hash, size, member count and fixed verification revision.
Use `tarfile.open(path, "r:")` and `extractfile(member)` to read a member without extracting paths. Four-camera packages contain `video_1..4.mp4`, `action_camera_1..4.json` and `systeminfo_1..4.json`. The action-camera files are **JSON arrays**, not JSON Lines. Row `frame` and `time` identify the time sample; camera position, quaternion and intrinsics retain the original conventions. Technical systeminfo files describe the capture surface, not a physical-device identity.
Historical SHA/QC files are source records, not a new certification. A `.public-copy.*` suffix identifies a small derivative with operator-private paths removed; its source digest is retained separately, and original attestations must not be applied to its changed bytes. No numeric measurements, failed-QC results or eligibility markers were upgraded.
## Review and limits
All structured JSON/JSONL metadata in the source scope was read, including all rows, and checked for account/contact/credential identifiers and operator-private paths. Archive directories and representative media/container metadata were inspected. Visual review samples frames; it is not exhaustive frame annotation or semantic audio review. Original media bytes are preserved and checked against available source SHA lists during transfer. Pinned Hugging Face LFS size/hash verification precedes source deletion.
This is an owner-authorized public project-closeout archive for inspecting capture quality, schemas and reproducibility. It is **not certified or newly authorized model-training data**. Original `non_trainable`, `internal-validation`, licence and QC statements are preserved; absent fields remain absent. Public access does not grant third-party game or asset rights. See `NOTICE.md` and the accompanying upstream notices.
Source and schema code: [CuddlerAI/Diogenes-Archive](https://github.com/CuddlerAI/Diogenes-Archive).