GraDBridge on M3FD

Official implementation of GraDBridge on the M3FD RGB-IR multispectral object detection dataset.

This repository provides a single-scale, single-fusion implementation of GraDBridge, in which the GNSB module is applied once for RGB-IR feature fusion. It is intended as a compact and reproducible version for training and evaluating GraDBridge on M3FD.

Overview

Figure 1. Overview of the GraDBridge framework.

GNSB Architecture

Figure 2. Detailed architecture of the fusion module used in the single-scale, single-fusion configuration.

Visualization

Figure 3. Qualitative detection results of GraDBridge on the M3FD dataset.

Repository Structure

GraDBridge/
├── configs/
│   └── hyp.scratch.yaml
├── data/
│   └── m3ddata.yaml
├── models/
│   ├── component/
│   │   └── GNSBOperatorBiasing.py
│   └── config/
│       └── GNSBOperatorBiasing_SingleFusion.yaml
├── train.py
├── test.py
└── train_m3fd.sh

Dataset

Configure the M3FD dataset paths and class information in:

data/m3ddata.yaml

The configuration should contain:

train_rgb:
val_rgb:
train_ir:
val_ir:
nc:
names:

The M3FD dataset itself is not included in this repository. Please specify the corresponding RGB and IR training/validation paths in data/m3ddata.yaml.

Installation

Install the required dependencies with:

pip install -r requirements.txt

Training

Training can be started directly with:

bash train_m3fd.sh

Alternatively, run the full command:

python train.py \
  --data data/m3ddata.yaml \
  --hyp configs/hyp.scratch.yaml \
  --cfg models/config/GNSBOperatorBiasing_SingleFusion.yaml \
  --weights yolov5l.pt \
  --batch-size 8 \
  --epochs 200 \
  --device 1