This repository provides a comprehensive index of open neuroimaging datasets specifically curated for reconstructing visual perception from human fMRI data. The guide targets AI and machine learning researchers who may lack familiarity with neuroimaging methodology, addressing common pitfalls in recent reconstruction approaches that have gained traction at major AI conferences.

The index includes datasets like vim-1, BRAINS, Miyawaki Dataset, BOLD5000, Generic Object Decoding, Natural Scenes Dataset, THINGS-fMRI, cNeuromod-THINGS, vim-2, and Doctor Who Dataset. Each dataset entry details visual field coverage, voxel size, fixation requirements, repetitions for signal-to-noise ratio, and copyright considerations. The repository distinguishes between identification, decoding, and true reconstruction tasks, emphasizing that genuine reconstruction requires generalization to novel stimuli beyond training categories.

Researchers are advised to evaluate datasets based on train-test independence, stimulus diversity, visual field coverage, and voxel resolution. The guide warns against common mistakes where models effectively perform classification within predefined candidate sets rather than true open-set reconstruction. Datasets with larger visual field coverage and higher voxel resolution from 7T scanners are preferred for capturing fine-grained visual information in early visual cortex.