KShivendu/abo-lens-embeddings

ABO product photos with SigLIP 2, EmbeddingGemma 2 and CLIP embeddings

61,105 products and 219,511 product photos from Amazon Berkeley Objects (ABO), each photo embedded with three models: SigLIP 2 base (768 dims), EmbeddingGemma 2 (768 dims, image token budget 70) and CLIP ViT-B/32 (512 dims). Built for Lens Search, a Google Lens style demo on Qdrant.

Tables

products: one row per product. product_id, item_id (Amazon ASIN), name, brand, color, material, style, product_type (ABO's type, for example SOFA), category (a coarser shop category), country, domain, photos.

photos: one row per photo. product_id, photo (path inside ABO's images/small/), url (public S3 link), is_main, shared_by, siglip2, embeddinggemma2, clip_vit_b32.

How it was made

• One product per main photo (ABO lists one product once per Amazon store), English names preferred. 2,000 of the 64,406 phone cases kept, so they do not crowd out everything else. Up to 4 photos per product.
• shared_by counts the products that use the exact same photo. A photo used by 3 or more products is a stock image (the same living-room shot sits in every AmazonBasics spotlight bulb listing). The Lens Search demo drops those from the index, except main photos.
• Vectors are L2-normalised float32.
• siglip2: google/siglip2-base-patch16-224 image features, fp32. For text queries, lower-case the text and pad to 64 tokens (padding="max_length", max_length=64). The browser build onnx-community/siglip2-base-patch16-224-ONNX gives the same vectors (median cosine 0.99 on 446 test photos).
• embeddinggemma2: photos take no task prefix. For text queries use task: search result | query: <your text>.
• clip_vit_b32: openai/clip-vit-base-patch32 image embeds.

Use with Qdrant

Each product as one point with a multivector (one vector per photo), scored by its best photo:

from qdrant_client import QdrantClient, models
client.create_collection("products", vectors_config={"image": models.VectorParams(
    size=768, distance=models.Distance.COSINE,
    multivector_config=models.MultiVectorConfig(comparator=models.MultiVectorComparator.MAX_SIM))})

The demo adds a second vector per product, the main photo alone, and ranks by 0.7 * main + 0.3 * best photo. With unseen photos of 446 products, searched among all products, the exact product lands in the top 10 for:

vectors	exact product in top 10
SigLIP 2 (live demo)	68.8%
EmbeddingGemma 2	69.5%
CLIP ViT-B/32	36.5%

License

ABO is licensed CC BY 4.0 by Amazon.com, Inc. (Collins et al., "ABO: Dataset and Benchmarks for Real-World 3D Object Understanding", CVPR 2022). The embeddings are derived from it under the same license.