Flash‑Splat Dataset Overview

Dataset Description

Flash‑Splat is a multi‑view flash/no‑flash dataset for 3D reflection removal. It includes images captured under flash and no‑flash conditions, along with corresponding camera poses and point‑cloud initializations produced by COLMAP.

Dataset Download

The dataset can be downloaded from: Google Drive

Contents

• Multi‑view flash/no‑flash images
• Camera poses
• Point‑cloud initialization

Usage

The dataset supports training and testing the Flash‑Splat model.

1. Download and extract the dataset to a chosen directory.
2. Run the following command for 3D reflection separation:

   python main.py --data_dir DATASET_DIRECTORY
   

3. To log training results with Weights & Biases, run:

   python main.py --data_dir DATASET_DIRECTORY --turn_on_wandb
   

Output

• GIF files of the separated transmission and reflection 3D scenes stored in demo/.
• Model checkpoints stored in experiments/.