SeeThroughSmoke Dataset

Paired surgical smoke / ground-truth image dataset for training and evaluating smoke removal (desmoking) models in laparoscopic surgery.

Paper: Seeing Through Smoke: Surgical Desmoking for Improved Visual Perception (arXiv:2603.25867)

Project Page: smoke.surgeryvision.org

Demo Videos

Liver — Ours (Base Model)	Liver — Ours (Fine-tuned)
Heart — Ours (Base Model)	Heart — Ours (Fine-tuned)

Supplementary Video

Dataset Structure

Each row in the viewer shows: Smoked Input (image) → Ground Truth (gt) for the given organ.

SeeThroughSmoke/
├── train/                    # 3,097 paired images
│   ├── metadata.csv
│   ├── liver_kidney/         # 1,282 pairs
│   │   ├── GT/
│   │   └── smoke/
│   ├── small_Intestine/      # 648 pairs
│   ├── large_intestine/      # 596 pairs
│   ├── unknown/              # 357 pairs
│   └── uterus/               # 214 pairs
└── test/                     # 2,720 paired images
    ├── metadata.csv
    ├── brain/                # 490 pairs
    ├── heart/                # 672 pairs
    └── stomach/              # 1,558 pairs

Statistics

Split	Organ	Pairs
Train	liver_kidney	1,282
Train	small_Intestine	648
Train	large_intestine	596
Train	unknown	357
Train	uterus	214
Train Total		3,097
Test	brain	490
Test	heart	672
Test	stomach	1,558
Test Total		2,720

Image Format

• Resolution: 1280 × 1024 pixels
• Format: PNG, 8-bit RGB
• Each smoke/ image has a corresponding GT/ image with the same filename

Naming Convention

Files follow the pattern: {organ}_{instance}_{side}_{frame}.png

• organ: anatomical region
• instance: patient/scan identifier
• side: left or right (camera orientation)
• frame: frame index within the sequence

Usage

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="artJiang20/SeeThroughSmoke",
    repo_type="dataset",
    local_dir="./SeeThroughSmoke"
)

Or use the hf CLI:

hf download artJiang20/SeeThroughSmoke --repo-type dataset