fedehorl/stanford-dogs-amplified

Stanford Dogs Amplified (Parquet Edition)

This dataset is an amplified and modernized version of the classic Stanford Dogs Dataset. It builds upon the original 120 dog breeds by automatically fetching, cleaning, and injecting thousands of new high-quality images scraped from Bing, filtered dynamically via YOLO object detection.

This specific repository hosts the dataset natively in Hugging Face's optimized Parquet format.

This means:

1. It is a strictly "Image Classification" dataset (the original XML bounding boxes have been removed).
2. It contains exactly two columns: image (the binary picture) and label (the breed integer id).
3. It has been pre-split into an 80/20 train/test stratified split, ensuring perfect class balance.
4. It can be downloaded and streamed instantly.

Quick Start

Load the dataset instantly into memory using the datasets library:

from datasets import load_dataset

# This will download the Parquet files and load the Train & Test splits
dataset = load_dataset("fedehorl/stanford-dogs-amplified")

# View the structure
print(dataset)
# DatasetDict({
#     train: Dataset({
#         features: ['image', 'label'],
#         num_rows: 34157
#     })
#     test: Dataset({
#         features: ['image', 'label'],
#         num_rows: 8540
#     })
# })

# Access an image and its label
sample_image = dataset['train'][0]['image']
sample_label = dataset['train'].features['label'].int2str(dataset['train'][0]['label'])
print(f"This is a {sample_label}")

Dataset Structure

• image: A PIL.Image.Image object containing the RGB image.
• label: A datasets.ClassLabel representing the dog breed (0 to 119).

Data Splits

The dataset contains a total of ~42,697 images natively split into:

• Train (80%): ~34,157 images
• Test (20%): ~8,540 images

Both splits are strictly stratified by the label column to maintain an identical distribution of the 120 dog breeds across the training and testing sets.

Dataset Creation & Augmentation Process

The augmented images were sourced by scraping Bing Images for modern pictures of the 120 breeds. To ensure dataset quality, a strict automated filtering pipeline was applied:

1. Format Constraints: Only strictly RGB .jpg files were allowed. Files with alpha channels, or in webp/avif formats, were dropped or converted.
2. YOLO Validation: Every scraped image was processed through the yolo26m model.
3. Confidence Threshold: Only images where YOLO successfully detected a "Dog" were admitted into the final dataset.