Smithsonian Butterfly Images with CLIP Similarity Scores

Encompassing a subset of 1,000 Smithsonian butterfly images processed for training a GAN model. It includes a 'sim_score' feature generated by a CLIP model based on descriptive text prompts. The data was curated by 'huggan' and last updated in April 2022.

Use Cases

• Train a GAN model using the 1,000 preprocessed butterfly images and their associated 'sim_score' for conditional generation.
• Analyze the distribution of 'sim_score' values across the image subset to understand feature similarity to text prompts.
• Use the background-removed images for computer vision tasks focused on butterfly morphology without environmental clutter.

Strengths

• Contains 1,000 curated images, providing a focused subset for model training.
• Includes a 'sim_score' feature derived from CLIP model evaluation against specific text prompts.
• Images have undergone preprocessing steps including background removal to isolate subjects.

Limitations

• The dataset is a subset limited to 1,000 images, which may be insufficient for training large-scale models.
• Preprocessing steps like background removal may have introduced artifacts or reduced visual quality for some images.
• The dataset's scope is limited to butterfly species from the Smithsonian collection, potentially lacking diversity.

Provenance

Source

Subset derived from the 'ceyda/smithsonian_butterflies' dataset on Hugging Face.

Collection Method

Images were processed by adding CLIP-based similarity scores, removing duplicate species, limiting the count, and removing backgrounds.

Freshness

Last updated on 2022-04-16.

Geography

Region tag indicates 'us', suggesting a United States focus, but specific geography is unknown.