ImageNet 1K Dataset with 256x256 Lanczos-Resampled Images

ImageNet 1K is a repackaged version of the ILSVRC 2012 dataset containing images organized according to the WordNet hierarchy. The images have been center-cropped to square and resampled to a uniform 256x256 pixel resolution using Lanczos resampling. This specific repack was created by benjamin-paine and uploaded to Hugging Face in September 2024.

Use Cases

• Train multi-class image classification models using the 1000 object categories defined by the WordNet hierarchy.
• Benchmark model performance on a standardized, preprocessed version of the ILSVRC 2012 validation set with 256x256 images.
• Use the center-cropped and Lanczos-resampled image features for transfer learning on downstream computer vision tasks.
• Analyze model robustness using the consistent image preprocessing applied across all samples in the dataset.

Strengths

• Contains the complete ILSVRC 2012 (ImageNet 1K) dataset, a foundational benchmark with over 1 million images.
• All images are preprocessed to a consistent 256x256 resolution using Lanczos resampling, reducing initial data preparation overhead.

Limitations

• The original ImageNet dataset has known geographic and demographic biases, primarily representing North American and Western contexts.
• The center-cropping transformation may remove contextual information from the edges of non-square original images.
• The dataset is a repack of the 2012 benchmark and does not include more recent data.

Provenance

Source

ILSVRC/imagenet-1k (ImageNet Large Scale Visual Recognition Challenge 2012).

Collection Method

Images were crowdsourced and organized according to the WordNet hierarchy. This repack applied center-cropping and Lanczos resampling to 256x256.

Time Range

null

Freshness

Last updated on the platform in September 2024, but the underlying data is from the ILSVRC 2012 competition.

Geography

Primarily United States, with known geographic bias.