LogiCo

Official implementation of "LogiCo: A Unified Framework for Logical and Structural Anomaly Detection (ECCV 2026)"

LogiCo is a unified framework for logical and structural anomaly detection that employs a novel component-level feature reconstruction technique to capture inter-component logical constraints. Specifically, LogiCo maps pre-trained image features into a discrete component-level feature space and performs collaborative feature reconstruction at both the component and patch levels, enabling effective detection of both logical and structural anomalies. Furthermore, to address the specific challenge of count-related logical anomalies, we integrate a segmentation-map discriminator that extends the model’s capability to identify quantitative inconsistencies. LogiCo achieves state-of-the-art performance on multiple logical and structural anomaly detection benchmarks.

🔧 Installation

To run experiments, first clone the repository and install requirements.txt.

$ git clone https://github.com/cnulab/LogiCo.git
$ cd LogiCo
$ pip install -r requirements.txt

Download Pre-trained Models

Request DINOv3 weights, and download DINOv3-B/16 (dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth), DINOv3-L/16 (dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth), and DINOv3-based dino.txt (dinov3_vitl16_dinotxt_vision_head_and_text_encoder-a442d8f5.pth). Place them under the ckpt directory.

Download Datasets

• MVTec-LOCO [Official]
• MVTec-AD [Official]
• VisA [Official]
• Real-IAD [Official]

Place them under the data folder and perform preprocessing. Please refer to data/README.

🚀 Experiments

Use the following command for component segmentation, taking Breakfast_box from MVTec-LOCO as an example:

$ python create_segmentation_maps.py --dataset mvtec_loco --category breakfast_box

You can skip this step, download the segmentation maps from the link below, and place them in the data folder:

• mvtec_loco segmentations
• mvtec segmentations
• visa segmentations
• real-iad segmentations

Train LogiCo using the following command:

$ python train_logico.py --dataset mvtec_loco --category breakfast_box  --save_dir saved_results

Evaluation and visualization:

$ python evaluation.py --dataset mvtec_loco --category breakfast_box  --save_dir saved_results

The complete running scripts are provided in the runs folder.

🔗 Citation

If this work is helpful to you, please cite it as:

@inproceedings{zhang2026logico,
      title={LogiCo: A Unified Framework for Logical and Structural Anomaly Detection}, 
      author={Ximiao Zhang, Min Xu, and Xiuzhuang Zhou},
      year={2026},
      eprint={2606.28688},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}