How Attribute‑Specific Embedding Networks Revolutionize Fashion Copyright Protection

Background and Motivation

When brand A releases an original dress, brand B often copies it by slightly modifying the length, collar, or sleeves to avoid copyright disputes. Traditional copyright detection compares whole images, which fails to catch such local modifications.

Proposed Method

Alibaba Security and Zhejiang University propose a fine‑grained similarity learning method that focuses on attribute‑specific regions. The method learns separate embedding spaces for each fashion attribute (e.g., collar shape, sleeve length) and measures similarity within those spaces, enabling accurate detection of “local plagiarism”.

Attribute‑Specific Embedding Network (ASEN)

The ASEN model consists of three components: a CNN feature extractor, an Attribute‑Specific Spatial Attention (ASA) module, and an Attribute‑Specific Channel Attention (ACA) module. ASA uses the given attribute to locate relevant garment parts, while ACA refines channel‑wise features that are also guided by the attribute. Both modules operate in parallel, allowing end‑to‑end learning without interference between different attribute similarity measures.

Experiments

ASEN was evaluated on FashionAI, DARN, DeepFashion, and Zappos50k datasets for attribute‑specific retrieval and triplet prediction. Quantitative results (Tables 1‑4) show consistent and significant improvements over baseline methods in mean average precision. Qualitative results demonstrate that the model correctly retrieves garments sharing the target attribute while varying other aspects such as style, color, and background.

Qualitative Visualizations

Visualization of the ASA module shows that the network can localize attribute‑related regions even under complex backgrounds and poses. Retrieval examples on FashionAI illustrate that the model retrieves items sharing the queried attribute (e.g., V‑neck, sleeveless) while presenting diverse styles, colors, and backgrounds.

Impact

By measuring fine‑grained similarity, ASEN can be used to re‑rank results of standard in‑shop clothing retrieval, improving the relevance of the final list. The approach also provides a powerful tool for fashion copyright enforcement.

Paper: https://arxiv.org/abs/2002.02814

References

Liu, Z. et al., DeepFashion: Powering robust clothes recognition and retrieval with rich annotations, CVPR 2016.

Ak, K. E. et al., Efficient multi‑attribute similarity learning towards attribute‑based fashion search, WACV 2018.

Huang, J. et al., Cross‑domain image retrieval with a dual attribute‑aware ranking network, ICCV 2015.

Ji, X. et al., Cross‑domain image retrieval with attention modeling, ACM Multimedia 2017.

He, R. et al., Learning compatibility across categories for heterogeneous item recommendation, ICDM 2016.

Vasileva, M. I. et al., Learning type‑aware embeddings for fashion compatibility, ECCV 2018.

Veit, A. et al., Conditional similarity networks, CVPR 2017.

Hu, J. et al., Squeeze‑and‑excitation networks, CVPR 2018.