[ICCV 2025] SALAD

Semantics-Aware Logical Anomaly Detection

Matic Fučka, Vitjan Zavrtanik, Danijel Skočaj

University of Ljubljana, Faculty of Computer and Information Science

Overview | Get Started | Results | Reference | Questions

Overview

This repository contains official PyTorch implementation for SALAD - Semantics-Aware Logical Anomaly Detection (Accepted to ICCV 2025).

SALAD (Semantics-Aware Logical Anomaly Detection) extends surface anomaly detection to capture both structural (scratches, dents) and logical (missing/misplaced parts) anomalies. First, we introduce a novel composition map generation process, that enables direct extraction of composition maps from the image without the need for hand labels or category-specific procedures.

Second, it introduces a compositional branch that learns the normality distribution of object composition maps. SALAD sets a new state-of-the-art on MVTec LOCO (96.1% AUROC).

Get Started

Environment setup

Create a Python environment and install required packages:

conda create -n SALAD python=3.10 pip
conda activate SALAD
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install opencv-python
pip install segment-anything
pip install segment-anything-hq
pip install scikit-learn
pip install timm
pip install positional-encodings[pytorch]
pip install imgaug
pip install numpy==1.26
pip install pandas

Download MVTec LOCO from this link or via the supplemented script:

./download_loco.sh

Composition map generation

Already generated composition maps can be downloaded via the following script:

./download_composition_maps.sh

In the case you wish to generate your own composition maps first generate background masks with SAM (download the pretrained weights using the script download_sam.sh) run:

python python create_fg_masks.py

Then create the pseudo labels with:

python python create_pseudo_labels.py --category <category_name>

In the end train a composition map segmentation model with:

python train_composition_segmentation_model.py --category <category_name>

Training and evaluation

Run the training and evaluation with:

python train_salad.py --category <category_name>

NOTE: ImageNet has to be downloaded to enable training of EfficientAD.

Run evaluation only:

python test_salad.py --category <category_name>

Pretrained weights and results

Download the pretrained weights using the following script:

./download_pretrained_weights.sh

and evaluate it with the evaluation command above.

Reference

If you found this work useful, consider citing our paper and giving this repo a ⭐ 😃

@InProceedings{fucka_salad,
  title={{SALAD} -- {S}emantics-{A}ware {L}ogical {A}nomaly {D}etection},
  author={Fu{\v{c}}ka, Matic and Zavrtanik, Vitjan and Sko{\v{c}}aj, Danijel},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2025},
  month={October}
}

Questions

For issues or questions, please open a GitHub issue or email the author directly.

Acknowledgement

Thanks to ComAD and EfficientAD for inspiration.