CryptoVision: Processed Lab Images of 81 Fish Species

7,944 processed image arrays of shape 299x299x3, derived from laboratory images of fish. The dataset, created by Leonardo Reginato, includes taxonomic labels and is split into 5,560 training, 1,192 validation, and 1,192 test samples. It is intended for reproducibility of the CryptoVision study and was last updated in May 2026.

Use Cases

• Training image classification models based on the 81 labeled fish species.
• Evaluating model performance using the predefined train/validation/test splits.
• Reproducing computational analyses from the CryptoVision study based on the provided processed arrays and source code.

Strengths

• Includes 7,944 processed samples with consistent 299x299x3 dimensions.
• Provides a clear data split of 5,560 training, 1,192 validation, and 1,192 test samples.
• Covers 81 distinct species, offering taxonomic diversity for classification tasks.
• Includes checksums and preprocessing settings for reproducibility.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset consists of processed, resized arrays, not the original raw laboratory images.

Provenance

Source

Fish & Functions Lab, Marine Science Institute, University of Texas at Austin.

Collection Method

Derived from laboratory image data, resized and processed into RGB arrays.

Freshness

Last updated 2026-05-15 21:14:02