SceneBench v1.0.0

2230 source-derived multiple-choice questions across 28 task types and six families, generated anew from nine held-out CholecT45 videos. One test benchmark; no training or validation split is distributed.

Family	Questions
Perception	334
Relation	480
Composition	506
Procedure	300
CVS	219
Dynamic	391

Browse individual examples

Open the Dataset Viewer to browse the 2,230 examples, including the rendered marker images, question, options and reference answer. Click a row to expand it; Dynamic examples contain two images in chronological order. The repository is public as of 2026-09-23.

Loading

from datasets import load_dataset
ds = load_dataset("EgoF0102/SceneBench", split="test")
row = ds[0]
images = row["images"]  # List of PIL images; one, or two in chronological order.
print(row["question"], row["options"], row["answer"])

Image bytes are embedded in Parquet. Yellow box A / arrow A is already rendered for P01/P02. Do not pass answer fields, provenance, video IDs, timestamps or source graphs to a direct-VQA baseline. Options are presented in stored order. P03/V04 have 2 options, V01–V03 have 3, all others have 4. The answer letter is determined by answer_index. Each question has one source-derived answer.

Sources and split

Test videos: VID02, VID06, VID14, VID23, VID25, VID50, VID51, VID66, VID79. All other videos must remain outside this test split when constructing training data. Indices are zero-based 1 fps indices; original Cholec80 frame index is 25 times this index.

• Cholec80 / EndoNet: phase and tool presence.
• CholecT45 / Rendezvous: images and instrument-verb-target annotations. Cite Nwoye et al., Rendezvous: Attention Mechanisms for the Recognition of Surgical Action Triplets in Endoscopic Videos, Medical Image Analysis, 2022.
• SSG-VQA: supplied scene graphs, object categories, boxes and spatial relations; cite Yuan et al., Advancing Surgical VQA with Scene Graph Knowledge, IJCARS 2024 (preprint 2023). See local source provenance for exact records.
• Cholec80-CVS: segmented CVS criterion scores. Follow the original paper and source repositories for full author attribution and citations.

Generation and quality

Queries are deterministic functions of source labels. Ambiguous action queries are rejected. Spatial directions use 430×240 bbox-center geometry, a 5% axis margin, and matching positive/inverse relation edges. Graph-detected tool categories are cross-checked against source tool-presence labels. Per-task answer/video balancing, spacing and image deduplication limit repetition; actual counts and distributions are in dataset_statistics.json. Source frames are reused at most twice, and Dynamic endpoints are exclusive to their one pair. Dynamic questions compare two endpoint states, not tracked physical instrument identities.

101 of 2331 selected candidates were removed during visual screening; 18 anatomy arrow tips were adjusted within the source box. P01/P02/P04/C06 received screening of all selected items; C04/C05/CVS received answer-stratified screening. This was Codex visual screening, not expert surgical adjudication. The source scene graphs contain model-generated detections and can contain residual errors. This release is source-consistent, not a claim of independently expert-verified gold for every image. Review scope and exclusions are public in curation_report.json. No baseline predictions were used for curation.

CVS interpretation

Only Calot triangle dissection before the first clipping/cutting transition is eligible. Annotated intervals include both endpoints. Uncovered segments inside the valid domain default to (0,0,0); labels are not forward-filled. Conflicting overlapping scores are excluded. V01–V03 use the native 0/1/2 criterion scores; V04 follows the source rule total score >=5. This is a dataset convention, not independent clinical certification. Multiple spaced/deduplicated frames per interval are allowed.

V04 has only three positive frames, all from one event in VID66 (661–675 seconds). V03 score 2 has the same event concentration. V02 score 2 occurs in two VID51 intervals. Multiple frames do not create independent clinical events. Some V04 negatives are matched to VID66 and late Calot. Results for rare criteria must be interpreted at video/event level.

Evaluation

The baseline protocol uses qwen/qwen3-vl-32b-instruct through OpenRouter, temperature 0, one call per question, only exported images + question/options, returning a single option letter. It is a zero-shot base-Instruct direct-VQA baseline, without SFT/RL or predicted-graph conditioning. Results and a separate Chinese analysis document are added after complete evaluation. Total micro accuracy and descriptive family/task/class breakdowns refer to this same benchmark.

License and reproducibility

CC BY-NC-SA 4.0. Images and source-derived annotations retain source attribution, non-commercial restrictions and share-alike terms. See the license and original source publications. Export processing adds markers, resizes longest edge to at most 1280 and encodes JPEG quality 95; no crop or flip.

Full Chinese task definitions, real examples, source mapping, sampling and limitations: Benchmark说明. Generator, fixed curation decisions and validator: code/. Credentials and raw unused source data are not included.

Frozen benchmark SHA256: 260d6a03c822f6d80e0eb4b585924e3d1910e39f0503ded056bbb61132276bc6.

Completed baseline results

Model qwen/qwen3-vl-32b-instruct: 1105/2230 = 49.55% strict micro accuracy; valid-answer coverage 100.00%. API errors: 0; parse errors: 0. Six-family equal mean: 48.50%.

Family	Correct / N	Accuracy
Perception	182/334	54.49%
Relation	302/480	62.92%
Composition	216/506	42.69%
Procedure	140/300	46.67%
CVS	82/219	37.44%
Dynamic	183/391	46.80%

Full results and limitations, machine-readable statistics, raw predictions.