z81980440/vault_dhaka_det_cvat

Dhaka Drone Vehicle Detection — CVAT Segmentation Datasets (vault)

Encrypted per-video CVAT (COCO 1.0) instance-segmentation datasets produced by the repeatable Dhaka drone vehicle-detection pipeline over the private input bucket hamimmahmud0/dhaka-drone-vehicle-detection.

Dataset table

Each zip contains dataset/<video>/images/<video>_f%06d.jpg (native 3840x2160) + dataset/<video>/annotations.json (COCO 1.0).

Decrypt + import instructions

Password: the shared vault password (zip password policy — see MANIFEST.txt run id task_dhaka-drone-vehicle-det-2026-09-04T181609Z; all tokens/passwords/ files are identical). Decrypt with ZipCrypto:

7z x -p"$PASS" cvat_DJI_0409.zip     # or: unzip -P "$PASS" cvat_DJI_0409.zip

Then import dataset/DJI_0409/annotations.json + the adjacent images/ folder into CVAT via Import dataset → COCO 1.0 (task with image data; RLE segmentation, 11 categories, ids 1–11).

Processing parameters (locked, identical for every video/run)

• —Sampling: 1 frame per 5 s (ffmpeg -vf fps=1/5 -q:v 2), includes t≈0; count = floor(duration/5) ±1.
• —Model: SAM3.1 Object Multiplex, native full-resolution full-frame (no grid).
• —Prompts (CVAT category ids 1–11, fixed order): 1 Vehicle, 2 Person, 3 car, 4 bus, 5 truck, 6 van, 7 rickshaw, 8 motorcycle, 9 bicycle, 10 cart, 11 pickup.
• —Confidence threshold: 0.25 per class.
• —Dedup (per frame, before CVAT): a generic Vehicle mask is dropped iff its MASK IoU vs any subclass mask (car…pickup) is ≥ 0.5; subclasses always kept; Person never participates; same-class near-duplicates (IoU ≥ 0.9) keep the higher score; non-overlapping Vehicle detections are kept.
• —Bounding boxes are the tight boxes of the segmentation masks (COCO convention); area = mask pixel area; counts are ASCII str RLE (CVAT-import safe).

Security note (accepted residual risk)

Archives are encrypted with ZipCrypto (7z -mem=ZipCrypto, level 1), which is a weak cipher (known-plaintext attackable) — accepted per locked policy for consistency with the DRINF vault practice. Do not rely on these archives for long-term secrecy.

Contents / manifest

• —cvat_DJI_0409.zip, cvat_DJI_0412.zip — encrypted per-video CVAT datasets.
• —MANIFEST.txt — append-only record (file, bytes, md5, sha256, params, run, date).
• —PROCESSED.json — machine-readable processed-set (watch manifest: match key = file name + md5).
• —dataset_grid_DJI_0409.png, dataset_grid_DJI_0412.png — plaintext QA montages (all 1/5 s thumbnails with mask overlays, frame captions).

Input videos remain in the private bucket hamimmahmud0/dhaka-drone-vehicle-detection (auth required). New videos uploaded there are processed by the same pipeline and appended to this vault. Source videos: DJI0409.MP4 (md5 e40b33171506c729276a2fa3fe296ff7), DJI0412.MP4 (md5 f2a09113abcf9a736044a3b197fb7c71).