HeyDonto/sd2-staged-foreign-objects

SD2 staged-laboratory foreign-object frames (heydonto)

97 frames · 163 frame-level annotation rows · 8 staged laboratory takes · CC BY 4.0

This dataset discloses and carries the SD2 own-footage portion of the training data of the ORena SAVE FOCUS challenge entry's FRAME-track component: 163 of that component's 58,086 pooled training rows. The other sources of that corpus are not part of this dataset.

What is in it

• frames/ — 97 JPEG frames (filename = first 16 hex of the frame's sha256 + take id + time in seconds), extracted from the platform encodes (mp4) of eight staged laboratory recordings (session staged_session2, recorded 2026-08-24/25; the AVI masters are retained privately and identified by sha256). The recordings are synthetic physical scenes (dataset family sd2_synthetic_physical): surgical props and butcher-meat tissue arranged on a locked setup. No patients are involved (program operator brief). Depiction: No individuals appear in the frames; the recording operator's thumb may appear occasionally when a QR code is shown (owner's statement, 2026-09-02, quoted verbatim below). The footage is the publisher's own, released under CC BY 4.0. Owner's statement, verbatim (2026-09-02 16:0xZ, owner's typed words, relayed by the program coordinator): "The recording does not show anyone, my thumb maybe sometimes when i show the QR. and does not show anyone else." (see CONSENT_ATTESTATION.json).
• sd2_annotations.parquet / sd2_annotations.jsonl — 163 rows in the challenge organizer's 14-column form (id, video, procedure_type, question, answer, answer_format, track, generation, clinical_relevance, ood, timestamp_start, timestamp_end, primary_capability, secondary_capabilities) plus provenance columns (frame_file, frame_sha256, t_seconds, take, proposal_source, reviewer_signature_id, signer_class, adoption_event_id, provenance_row_id, license).
• CONSENT_ATTESTATION.json, LICENSE, MANIFEST.json (sha256 and byte size of every file).

Answer formats: {"yes_no": 66, "state_label": 97}. Rows per take: {"D2-DRN01": 41, "D2-SPEC02": 15, "D2-ST01": 33, "D2-ST02": 18, "D2-ST03": 23, "D2-ST04": 12, "D2-ST06": 8, "D2-ST06R": 13}. Frames per take: {"D2-DRN01": 22, "D2-SPEC02": 10, "D2-ST01": 19, "D2-ST02": 11, "D2-ST03": 15, "D2-ST04": 7, "D2-ST06": 5, "D2-ST06R": 8}.

Provenance

Labels are machine state spans with human-verified labels (signer_class = machine_lineage_human_reviewed; signer trainer_study:machine-state-spans+desk-verified-labels): the label origin is machine/analysis-derived (the trainer-study intake ledger's state spans, an analysis converged with the owner using a general-purpose VLM), and every label was verified by the annotation desk's frame reads before adoption — they are not human-authored annotations. Adoption event ADOPT-2026-08-26-SD2ADOPTION2 (state/temporal + instrument/QC only). timestamp_start/end are floor(t_seconds) as HH:MM:SS; t_seconds is exact. The source recordings (AVI masters, ~3.3 GB) are not part of this dataset; they are retained privately as evidence and are identified in the program's records by sha256.

Intended use

Research on foreign-object recognition and state description in laparoscopic-style staged scenes; this is the SD2 own-footage part of the ORena entry's FRAME-track training corpus. Frame-level questions only. Load the annotations from sd2_annotations.parquet (the README configs pins it as the default data file; the .jsonl mirror is value-identical but its all-empty secondary_capabilities lists infer as null-typed under generic JSON loaders).

License and attribution

CC BY 4.0. Please attribute: heydonto — SD2 staged-laboratory foreign-object frames (2026).