FIND: Few-Shot Anomaly Inspection (ICCV 2025)

Unofficial PyTorch reimplementation of the paper:
FIND: Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data
📄 Paper

Note: No official code has been released by the authors.
v2-paper-aligned branch is the most up-to-date implementation.
This is an independent reimplementation based on the paper.

Few-shot support: K=5, 10, 50 supported via K_SHOT config.
Full-shot results reported. Few-shot benchmark results coming soon..

---

⭐ If you find this useful

If this reimplementation helped your research or learning, please consider starring the repo — it helps others find it and motivates further work!

---

Overview

This repository provides a code reimplementation of FIND for few-shot anomaly inspection using normal-only multi-modal data (e.g., RGB and surface normals).

Reproduced Results (Dowel, Full-shot, Single Run)

Metric	Ours	Paper
I-AUROC	0.8746	0.979
P-AUROC	0.9872	0.995
AUPRO@30%	0.9732	0.986
AUPRO@1%	0.6136	0.982

Single run, single category (dowel, full-shot).
Full results across all 10 categories coming soon.

Requirements

pip install torch torchvision timm tifffile open3d tqdm scikit-learn opencv-python

Dataset

Download MVTec 3D-AD from here.

Training

Set CATEGORY in find_train.py then run:

python find_train.py

Evaluation

python find_eval.py
python evaluate_experiment.py

File Attribution

File	Source
find_train.py	Our reimplementation of FIND (ICCV 2025) — training pipeline
find_eval.py	Our reimplementation of FIND (ICCV 2025) — inference & map saving
evaluate_experiment.py	Official MVTec 3D-AD evaluation scripts (modified)
generic_util.py	Official MVTec 3D-AD evaluation scripts
pro_curve_util.py	Official MVTec 3D-AD evaluation scripts
roc_curve_util.py	Official MVTec 3D-AD evaluation scripts
lifind.yml	Conda environment for training (CPU)
lifindgpu.yml	Conda environment for training (GPU)

Attribution

MVTec 3D-AD evaluation scripts are from the official MVTec 3D-AD dataset page.

Visualization Examples (Dowel category)

See visualizations/ for full results (RGB input, surface normal, anomaly map, overlay, GT mask).