Shopformer: Transformer-Based Framework for Detecting Shoplifting via Human Pose

Overview

This repository contains the official implementation of Shopformer (CVPR 2025). Shopformer is a novel transformer-based framework designed for detecting shoplifting behaviors using only human pose data, , rather than raw pixel information. Unlike traditional video-based methods, Shopformer focuses on privacy-preserving, pose-based shoplifting detection. The model introduces a two-stage architecture: A Graph Convolutional Autoencoder (GCAE) learns rich spatio-temporal embeddings from human pose sequences. These embeddings are tokenized and passed through a transformer encoder-decoder, which reconstructs the sequence. The reconstruction error is then used to compute a normality score for shoplifting detection.

Key Features

• GCAE-based tokenization of human pose sequences
• Transformer encoder-decoder with attention for behavior modeling
• Evaluated on PoseLift dataset (real-world shoplifting pose data)
• Privacy-preserving and real-time capable

Shopformer Architecture

The following figure illustrates the overall architecture of the Shopformer model:

Project Structure

• models/ – GCAE tokenizer & transformer model
• scripts/ – training and evaluation scripts
• data/ – instructions and expected format for PoseLift dataset
• config/ – training configurations
• utils/ – metric calculations, pose preprocessing

Dataset

This model is trained on the PoseLift dataset. You can access to the dataset nad related documentation here: PoseLift GitHub Repository

After downloading the dataset, organize the files into the following directory structure:

DATA/
└── Poselift/
   ├── gt/
   │   └── test_frame_mask/
   │       └── (test set frame-level binary mask files indicating normal or shoplifting behavior)
   └── pose/
       ├── train/
       │   └── (training pose JSON files)
       └── test/
           └── (test pose JSON files)

Installation

conda env create -f environment.yml

conda activate Shopformer

Training

Stage 1: Train the tokenizer using the following command, or use the pre-trained tokenizer

Stage 2: Freeze the encoder and train transformer: python3 main_to.py --dataset Poselift --model_optimizer adam --mask_root ...
--seg_len 24 --seg_stride 12 --num_kp 18 --model_num_heads 12 --model_num_layers 4
--epochs 10 --dropout 0.1 --model_lr 5e-05 --model_save_dir ... --model_loss mse --token_config graph --model_latent_dim 64

Results

Shopformer generates 2 tokens per pose sequence, as this setup achieved the best trade-off between accuracy and computational efficiency during ablation studies. Each token has an embedding size of 144, encoded using 8 channels over 18 keypoints. For detailed results comparing token counts ranging from 2 to 12, please refer to the ablation study section in the paper.

_{Table 1: AUC-ROC, AUC-PR, and EER of Shopformer compared with state-of-the-art pose-based anomaly detection models on the PoseLift dataset.}

Methods	AUC-ROC	AUC-PR	EER
STG-NF	67.46	84.06	0.39
TSGAD	63.35	39.31	0.41
GEPC	60.61	50.38	0.38
Shopformer	69.15	44.49	0.38

Citation

If you find our work useful, please consider citing:

@article{rashvand2025shopformer,
  title={Shopformer: Transformer-Based Framework for Detecting Shoplifting via Human Pose},
  author={Rashvand, Narges and Noghre, Ghazal Alinezhad and Pazho, Armin Danesh and Ardabili, Babak Rahimi and Tabkhi, Hamed},
  journal={arXiv preprint arXiv:2504.19970},
  year={2025}
}

Contact

If you have any questions or need assistance, please contact the authors at nrashvan@charlotte.edu.