Potestates/sar-rarp50-focus-qa

SAR-RARP50 → FOCUS QA (derived annotations)

1 616 question–answer rows derived from the human-annotated segmentation masks of SAR-RARP50 (EndoVis 2022, robot-assisted radical prostatectomy), reformatted into the ORena FOCUS FRAME answer schema.

Released to satisfy the ORena SAVE FOCUS Challenge requirement that additional annotations created on public datasets be made publicly available with the submission.

What this is — and is not

This is a narrow correction set built for one purpose, not a general-purpose surgical VQA resource. Its label distribution is deliberately degenerate: every fo_class row answers Needle and every binary row answers yes. That is what makes it effective as a 100-step corrective fine-tune, and also what limits its usefulness outside that role. It is published for transparency and reproducibility, not as a benchmark.

No images are redistributed. Rows reference the source video and frame index; obtain the frames from SAR-RARP50 directly.

Construction

Fully deterministic, from the dataset's own masks — no model-generated labels:

step	frames
annotated frames in the 3 videos available to us (video_01, video_02, video_04)	1 024
− dropped: catheter present (ambiguous against the FOCUS class External drain)	−469
− dropped: no suturing needle present	−151
retained, all needle-positive	404

Four fixed questions were instantiated per retained frame → 404 × 4 = 1 616 rows.

answer format	questions per frame	rows	answer
fo_class	3	1 212	Needle
binary	1	404	yes

Class mapping. suturing needle → Needle; thread, clamps, needle_holder, suction are instruments with no FOCUS foreign-object class and are ignored; catheter is ambiguous against External drain, so those frames are dropped entirely.

Frames were resized to 960 × 540 when used for training, matching the dominant FOCUS resolution — SAR-RARP50 is native 1080p, and leaving it would supply a resolution cue identifying the domain.

Two deliberate omissions

• No none-answer rows. none is never a correct fo_class answer in FOCUS (0 of 8 971 rows across train and test) despite 4 604 questions being phrased "…or answer with none". Needle-negative frames were therefore discarded rather than labelled none.
• No counting rows. The target model already under-counts; adding low-count examples from a sparse domain would have worsened that compression.

Construction script: build_sar_train.py in this repository.

Why it exists

The model being corrected reported surgical clips unconditionally on out-of-distribution anatomy: on prostate frames, questions whose answer contained "clip" were 0 % correct (n = 99), while on cholecystectomy they were 78.5 % correct (n = 261). Suppressing the class globally is harmful; the fix has to be conditional, and conditioning requires evidence from a domain where the class is genuinely absent.

Used for 100 steps at lr 2e-5, mixed at ~11–16 % of the training stream, this moved the official object_recognition_ood bucket by +6.58 points (45.18 → 51.76) at a cost of 2.62 points on the in-distribution bucket, for a net +1.14 overall.

Fields

video, timestamp_start, timestamp_end, id, question, answer, answer_format, primary_capability, procedure_type, track, generation, clinical_relevance, ood, secondary_capabilities.

Links

• Model: https://huggingface.co/Potestates/orena-focus-frame-qwen3.6-27b-lora
• Code and method description: https://github.com/DingTianxingjian/orena-focus-frame
• Source dataset: SAR-RARP50, EndoVis 2022

Licence

CC BY-NC-SA 4.0, inheriting the terms of SAR-RARP50. Only derived question–answer pairs and frame references are distributed here; the underlying video frames are not.