ImageNet Sketch Dataset with 50 Images per Class

Filled with 50,000 images, with 50 images for each of the 1,000 ImageNet classes. It was constructed by querying Google Images for sketches of each class name and manually cleaning the results.

Use Cases

• Test model generalization by evaluating a classifier trained on standard ImageNet against sketch-based images for the same 1,000 classes.
• Train a domain adaptation model to bridge the gap between photographic and sketch-based visual representations of the same object categories.
• Benchmark the robustness of computer vision models to stylistic variations, specifically black-and-white line drawings, across all ImageNet classes.
• Augment training data for sketch recognition tasks by leveraging the 50 images per class as a labeled sketch corpus.

Strengths

• Contains 50,000 total images, providing a substantial test set.
• Covers all 1,000 ImageNet classes, enabling comprehensive class-level analysis.
• Images were manually cleaned to remove irrelevant or misclassified results.

Limitations

• Images are limited to black-and-white sketches, lacking color and photographic texture.
• For some classes, the original query yielded fewer than 50 images, requiring augmentation via flipping and rotation, which may reduce diversity.
• The dataset is a static benchmark derived from web queries and may not represent the full stylistic variety of human sketches.

Provenance

Source

Constructed via Google Image queries for 'sketch of __' for each ImageNet class.

Collection Method

Images were queried, manually cleaned to remove irrelevant ones, and augmented with flips and rotations where necessary to reach 50 images per class.

Freshness

Last updated on 2024-09-10.