chadhas's picture
Upload folder using huggingface_hub
52a6d6f verified
|
Raw History Blame Contribute Delete
5.19 kB
metadata
license: cc-by-4.0
language:
  - en
pretty_name: Amazon Berkeley Objects products with clip-ViT-B-32 image embeddings
task_categories:
  - image-feature-extraction
  - feature-extraction
size_categories:
  - 10K<n<100K
tags:
  - qdrant
  - multimodal
  - e-commerce

amazon-berkeley-objects-clip-ViT-B-32-embeddings

This dataset contains 44,932 products from the Amazon Berkeley Objects (ABO) dataset, each with a CLIP embedding of its main image. Because CLIP maps text and images into the same space, you can find products by describing them in words or by showing a similar image.

The collection keeps ABO products that have an English name and a photo as their main image. Products that share a main image, usually size or color variants, are kept once, under the shortest of their names. Phone cases make up half of the source catalog, so 1,000 of them are kept, chosen at random.

Column Type Description
id string Point ID in the Qdrant collection
item_id string ABO item ID
name string Product name in English
brand string Brand, empty for about 7% of products
color string Color, empty for about 43% of products
product_type string ABO product type, such as CHAIR
country string Marketplace of the listing, such as US
image_url string The main image, at most 256 pixels on the longest side
vector list of float32 clip-ViT-B-32 image embedding, 512 dimensions, cosine

Generation process

The embeddings were generated with the clip-ViT-B-32 model from the image at image_url. Encode text queries with the same model to search the images:

from PIL import Image
from sentence_transformers import SentenceTransformer

image_path = "f3de0cfd.jpg"

model = SentenceTransformer("clip-ViT-B-32")
embedding = model.encode(Image.open(image_path))
query_embedding = model.encode("a red travel case")

On Qdrant Cloud, the same models are available through Cloud Inference as qdrant/clip-vit-b-32-vision and qdrant/clip-vit-b-32-text.

Using the dataset with Qdrant

The quickest way is to restore the ready-made snapshot, which includes the payload indexes. It requires Qdrant 1.20 or later:

PUT /collections/amazon-berkeley-objects/snapshots/recover
{
  "location": "https://snapshots.qdrant.io/amazon-berkeley-objects.snapshot"
}

To load the rows yourself, create the collection with a single unnamed vector, upload the rows, then create the payload indexes. The full-text index on name uses BM25 scoring, which also requires Qdrant 1.20 or later:

from datasets import load_dataset
from qdrant_client import QdrantClient, models

client = QdrantClient(url="http://localhost:6333")
client.create_collection(
    "amazon-berkeley-objects",
    vectors_config=models.VectorParams(size=512, distance=models.Distance.COSINE),
)

dataset = load_dataset(
    "Qdrant/amazon-berkeley-objects-clip-ViT-B-32-embeddings", split="train"
)
for batch in dataset.iter(batch_size=256):
    payload_fields = [name for name in batch if name not in ("id", "vector")]
    client.upsert(
        "amazon-berkeley-objects",
        points=[
            models.PointStruct(
                id=batch["id"][i],
                vector=batch["vector"][i],
                payload={
                    name: batch[name][i]
                    for name in payload_fields
                    if batch[name][i] is not None
                },
            )
            for i in range(len(batch["id"]))
        ],
    )

for field in ("item_id", "brand", "color", "product_type", "country"):
    client.create_payload_index(
        "amazon-berkeley-objects", field, models.PayloadSchemaType.KEYWORD
    )
client.create_payload_index(
    "amazon-berkeley-objects",
    "name",
    models.TextIndexParams(
        type=models.TextIndexType.TEXT,
        tokenizer=models.TokenizerType.WORD,
        lowercase=True,
        stemmer=models.SnowballParams(
            type=models.Snowball.SNOWBALL, language=models.SnowballLanguage.ENGLISH
        ),
        stopwords=models.Language.ENGLISH,
        scoring=models.TextScoringParams(type=models.TextScoringType.BM25),
    ),
)

License and attribution

ABO is licensed under CC BY 4.0. Credit for the data, including the images, goes to Amazon.com. Credit for building the dataset goes to its Amazon and UC Berkeley authors. This dataset is a modified subset of the ABO listings, as described above, with a CLIP embedding of each main image added.

Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F. Yago Vicente, Thomas Dideriksen, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. 2022. ABO: Dataset and Benchmarks for Real-World 3D Object Understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).