SmokeBench-SurgiMist

This is an anonymous review release for a staged surgical smoke restoration benchmark. It provides the primary B4 paired image resource, benchmark protocol, frozen split manifests, synthetic smoke generation parameters, external evaluation indices, real-smoke no-reference indices, code, result tables, and Croissant metadata.

The primary image release contains 8,000 Cholec80-derived paired examples: one synthetic degraded smoke image and one corresponding clean GT frame per record. The official paired benchmark split remains 5,600 train / 800 validation / 600 test records; the additional 1,000 paired images from videos 71-80 are included as reserved_source_holdout so the full 8k resource is available without changing the benchmark definition.

Why Hugging Face

We recommend hosting this anonymous package on Hugging Face Datasets. It supports arbitrary files, dataset cards, organizations, public dataset URLs, and Croissant export. The suggested anonymous URL placeholder is:

https://huggingface.co/datasets/anon-surgimist26/smokebench-surgimist

Create this under a fresh anonymous account or organization, not a personal or lab account, so contributor identity and commit metadata do not break double-blind review.

Contents

File	Records	Description
images/b4_full_8000/	8,000 pairs	Primary B4 paired image release with degraded/ and gt/ folders.
data/b4_full_8000_pairs.jsonl	8,000	Image-level pair manifest for the released degraded/GT images.
data/synthetic_benchmark_records.jsonl	35,000	B0-B4 video-disjoint synthetic benchmark records with generation parameters.
data/source_scaled_tier_records.jsonl	183,000	Cholec80 source-frame sampling tiers for restoration scaling experiments.
data/external_benchmark_records.jsonl	4,032	Manifest-only LSD3K and DesmokeData evaluation records.
data/real_smoke_records.jsonl	1,033	Manifest-only Cholec80 real-smoke and DeSmoke-LAP no-reference evaluation records.
metadata/	-	Split summaries, per-variant manifests, video lists, and release manifest.
code/	-	Smoke generation, split freezing, materialization, evaluation, plotting scripts, and a standalone smoke-synthesis code package.
results/	-	CSV summary tables used for the paper experiments.
croissant.json	-	Croissant metadata with core and minimal Responsible AI fields.

Benchmark Variants

The staged internal benchmark uses video-disjoint synthetic splits with Cholec80 videos 71-80 reserved for real-smoke evaluation:

Variant	Tag	Train	Val	Test
B0	depth_only_homogeneous	5,600	800	600
B1	single_layer_heterogeneous	5,600	800	600
B2	layered_static_no_optics	5,600	800	600
B3	layered_ns_no_optics	5,600	800	600
B4	ours_on_the_fly_ns	5,600	800	600

The released B4 full-8k image resource is organized as:

Split	Pairs	Role
train	5,600	Official paired training split.
val	800	Official paired validation split.
test	600	Official paired synthetic test split.
reserved_source_holdout	1,000	Extra paired source-holdout images from videos 71-80; not part of the official paired benchmark test.

Reproduction Scope

The main B4 degraded/GT image pairs are directly included under images/b4_full_8000/. To regenerate B0-B3 variants, source-scaling tiers, depth maps, or smoke-bank intermediates, obtain the original source datasets under their own terms, then use the provided split records and scripts to reconstruct the local experiment workspace.

The included scripts are the generation/evaluation entry points used by the project. Some materialization scripts expect local manifests with real clean_image and smoky_image paths after the source datasets have been arranged locally, so they should be run after reconstructing or regenerating those local manifests from the sanitized records.

For external benchmarks, use the manifest IDs to locate files in the original LSD3K, DesmokeData, and DeSmoke-LAP releases. This package does not grant rights to redistribute or use those datasets beyond their original licenses.

Croissant

croissant.json contains core Croissant metadata and minimal Responsible AI fields for NeurIPS E&D review. After uploading this folder to Hugging Face, also download the platform-generated Croissant file from the HF Croissant tab if available. If the generated file lacks RAI fields, submit this completed croissant.json to OpenReview or merge its RAI fields into the generated file before validation.

Limitations

The package is designed for anonymous review, benchmark evaluation, and non-commercial research. It is not a clinical tool and must not be used for patient-care decisions.

License and Attribution

Cholec80 is released by CAMMA under CC BY-NC-SA 4.0. The clean GT frames are sampled from Cholec80, and the degraded frames are adapted synthetic-smoke derivatives of those frames; therefore this image release is also distributed under CC BY-NC-SA 4.0. Users must keep attribution, indicate modifications, restrict use to non-commercial purposes, and share adapted material under the same license.

Please cite the Cholec80 / EndoNet publication when using the underlying data:

@article{twinanda2016endonet,
  title={EndoNet: A Deep Architecture for Recognition Tasks on Laparoscopic Videos},
  author={Twinanda, Andru P. and Shehata, Sherif and Mutter, Didier and Marescaux, Jacques and de Mathelin, Michel and Padoy, Nicolas},
  journal={IEEE Transactions on Medical Imaging},
  year={2016}
}