Oratis's picture
完成CARLA九组历史迭代归档并补齐全部文件和分片实际校验值
18ac886 verified
|
Raw History Blame Contribute Delete
5.02 kB
metadata
pretty_name: Diogenes CARLA Capture Iterations Archive v01
license: other
license_name: attributed-simulator-capture-archive
license_link: >-
  https://huggingface.co/datasets/DiogenesLab/Diogenes_CARLA_Capture_Iterations_Archive_v01/blob/main/NOTICE.md
tags:
  - video
  - simulation
  - robotics
  - multi-view
  - archival
size_categories:
  - n<1K

Diogenes CARLA Capture Iterations Archive v01

Historical CARLA camera-capture iterations: 260 MP4 files, 755 JSON metadata files, 187 PNG inspection frames/contact sheets and 20 empty historical status markers. The archive preserves 1,222 original objects / 24,090,799,445 bytes, across nine collections. Versions can overlap; video-file counts do not imply independent scenes or unique runs.

Collection MP4 files All objects Bytes
carla-final 24 108 2,196,036,361
carla-hz 24 132 2,293,821,926
carla-hz2 16 88 1,512,393,351
carla-road 20 112 1,903,560,577
carla-v1 40 130 3,557,444,427
carla-v2 40 170 3,750,163,949
carla-v3 40 160 3,667,947,562
carla-v5 28 161 2,630,682,431
carla-v6 28 161 2,578,748,861

Layout and interpretation

Tar shards preserve relative paths under collections/<collection>/. Video groups generally have four camera lanes video_1.mp4 through video_4.mp4, alongside action_camera_N.json, systeminfo_N.json, planning/provenance and available co-visibility/dynamic-object metrics. PNGs under sheets, ov or similarly named paths are original inspection views, not extra camera recordings. Some collections place additional planning files in subdirectories; the member manifest gives the exact layout.

Per-frame arrays contain time/frame/fps, virtual camera positions, rotations, intrinsics and speed, with version-dependent simulated player/input fields. The historical key camera_Follow Offset is intentionally unchanged. Some time fields are date-time strings; others are numeric seconds serialized as strings. Read each file's field_conventions: early planned metadata and later CARLA-native playback do not necessarily share an axis/rotation convention. Do not infer real human actions from synthesized input fields.

All representative videos are 60 seconds at nominal 30 fps. Earlier 4:3 views (1440×1080) coexist with 1920×1080 views. Existing media are not resized, trimmed or re-encoded. Some earlier videos have no audio; matching project tooling uses synthetic brown-noise ambience for muxed CARLA audio, not captured game-world sound. Treat that audio as synthetic and preserve the original per-file stream differences.

import json, tarfile
with tarfile.open("downloaded-shard.tar", "r:") as archive:
    for member in archive:
        if member.isfile() and member.name.endswith("provenance.json"):
            metadata = json.load(archive.extractfile(member))
            print(member.name, metadata.get("scene"), metadata.get("field_conventions"))

Quality and review

This archive intentionally includes failed and intermediate experiments: obstructed camera views, extreme or incorrect orientation, near-object clipping and differences in dynamic-object visibility. Original self-checks and source notes remain intact. The 20 empty .state-QC_PASS files are historical automation markers; they are not a new production certification or blanket evidence that all delivery requirements passed. Original unresolved coordinate-convention caveats and synthetic-scene flags remain unchanged. No train/test split, uniform action coverage or new training eligibility is claimed.

All 755 JSON objects were fully parsed and scanned. Sensitive-field review resolved 30 attribution hits as the upstream CARLA attribution text, which is preserved. Exact-SHA comparison reused the already fully reviewed camera-plan payloads where applicable; additional variant field names/types and timestamp formats were inspected from their originals. Representative media inspection covered 72 video first frames and 14 PNG frames/contact sheets across all nine collections. No private account or desktop overlay was identified in those samples. This is not exhaustive video viewing or audio transcription.

One 6,148-byte macOS .DS_Store record is retained privately and excluded from public capture data. Every other listed capture object, including empty status markers and defective iterations, is preserved byte-for-byte. Publication is complete: all 1,222 original capture objects are preserved in 26 TAR shards. MANIFEST.jsonl gives every member's actual size, SHA-256 and containing archive; ARCHIVE_INDEX.jsonl gives each shard's immutable upload revision, size and SHA-256. All shards have been verified against pinned Hugging Face LFS and anonymous range downloads before matching source removal. Source tools and known historical caveats: Diogenes Archive.