ImageNet 64x64 Downscaled Image Classification Dataset

A 64x64 pixel downscaled version of the ImageNet dataset, containing labeled images across 1,000 object categories. The dataset was created by a student and published on Hugging Face in April 2022. It is derived from the original ImageNet collection, intended for efficient model training and experimentation.

Use Cases

• Train image classification models to predict the 1,000 ImageNet class labels from 64x64 pixel RGB images.
• Benchmark model performance and efficiency on a standardized, lower-resolution version of the ImageNet challenge.
• Develop and test data augmentation or super-resolution techniques using the downscaled 64x64 images as inputs.
• Analyze the impact of image resolution reduction on feature extraction and model accuracy across different object categories.

Strengths

• Derived from the widely recognized and standardized ImageNet benchmark with 1,000 object classes.
• 64x64 pixel resolution reduces storage and computational requirements for faster experimentation.

Limitations

• Lower 64x64 resolution may lose fine-grained visual details present in the original high-resolution images, potentially limiting model accuracy.
• Specific sample count, class distribution balance, and preprocessing details are not provided in the metadata.

Provenance

Source

Derived from the ImageNet dataset.

Collection Method

Downscaling of original ImageNet images to 64x64 pixel resolution.

Time Range

null

Freshness

Last updated in April 2022.

Geography

null