FBR (Fuse Before Reconstruct)

[Information Fusion 2026] Fuse Before Reconstruct: A Novel Reconstruction Paradigm for Multimodal Industrial Anomaly Detection

by Zhi Li, Hanyu Wang, Qiyu Chen, Fei Shen, Zhengtao Zhang.

🛠️ Environment Setup

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
python -m pip install -e . --no-deps

pointnet2_ops contains CUDA extensions and must match the active PyTorch and CUDA versions.

📦 Pretrained Feature Extractors

FBR uses the same pretrained feature extractors as M3DM and CFM:

The Point-MAE checkpoint above is distributed by M3DM. Download it and place it at:

checkpoints/feature_extractors/pointmae_pretrain.pth

🗂️ Data Preparation and Preprocessing

Download MVTec 3D-AD and Eyecandies from their official websites:

datasets/
├── mvtec3d/
├── eyecandies/
└── eyecandies_preprocessed/

Preprocess MVTec 3D-AD:

python tools/preprocess_mvtec.py datasets/mvtec3d

This operation modifies the TIFF point-cloud files in place. Keep an untouched copy of the dataset.

Preprocess Eyecandies:

python tools/preprocess_eyecandies.py \
    --dataset_path datasets/eyecandies \
    --target_dir datasets/eyecandies_preprocessed

The target directory must not already exist.

🚀 Training and Evaluation

Dataset-specific files extend configs/base.yaml. The base configuration contains the model dimensions, backbone settings, 300-epoch training schedule, batch size 2, learning rate 1e-3, training seed 115, and evaluation seed 42.

Paths and experimental parameters can be changed in YAML or overridden from the command line:

fbr train \
    --config configs/mvtec.yaml \
    --classes bagel cookie \
    --epochs 300 \
    --bat