BrightCut: Localizing Visual Splices

Project Overview

BrightCut is an end-to-end research toolkit for identifying tampered regions within composite imagery. It couples convolutional backbones with refinement heads to deliver high-resolution localization masks that highlight suspected edits.

Feature Highlights

• Joint global and local reasoning for precise mask generation.
• Modular data preparation scripts supporting patch-based and full-resolution workflows.
• Checkpointed training utilities with resume-friendly logging.
• Visualization helpers for side-by-side label, prediction, and ground-truth comparisons.

Quickstart

pip install -r requirements.txt
python hybird.py --help

The first command installs dependencies. The second displays all configurable hyperparameters exposed by the training driver.

Dataset Preparation

This project supports the Columbia splicing dataset, available through the official request form. After obtaining the data, generate training patches and resized images:

python tools/make_dataset_colmbia.py /path/to/dataset

Adjust output locations inside the script or via command-line arguments to match your storage layout. The resulting structure separates training, validation, and testing splits for both patches and full images.

Training Workflow

Use train_local.sh as a template for launching experiments. Key flags include --epochs, --arch, --train-batch, --data, and --base-dir. Update paths before execution to point at your prepared dataset. Models and logs are stored under checkpoint/local by default.

Monitoring Progress

Training produces scalar summaries and segmentation previews. Point your preferred visualization tool at the checkpoint directory to monitor loss curves, accuracy trends, and qualitative mask outputs over time.

Evaluation and Results

Sample outputs provided below demonstrate label predictions, predicted masks, and ground-truth annotations. Use scripts/utils/evaluation.py to compute metrics such as pixel accuracy and mean IoU on your validation split.

Validation statistics including loss curves and segmentation accuracy are logged automatically during training.

Credits

This repository consolidates research code exploring semi-global strategies for splice localization. Cite the project if it aids your work and share improvements via pull requests.