This is the pre-processed fMRI data and frames sampled from videos in public cc2017 dataset[1], which is used by [2].
-subj01~3_train/test_fmri.pt: The significant voxels(Bonferroni correction, P < 0.05) were considered to be stimulus-activated voxels and used for subsequent analysis.
The following is a comparison of the number of voxel choices：
| Method | Subject 1 | Subject 2 | Subject 3 |
|---|---|---|---|
| MinD-Video | 6016 | 6224 | 3744 |
| NeuroClips | 13447 | 14828 | 9114 |
-GT_train/test_3fps.pt: The videos from the cc2017 dataset were downsampled from 30FPS to 3FPS to make a fair comparison with the previous methods.
-GT_train/test_caption/emb.pt: The pre-processed captions and their CLIP embeddings from BLIP-2.
Reference:
[1] Wen, Haiguang, et al. "Neural encoding and decoding with deep learning for dynamic natural vision." Cerebral cortex 28.12 (2018): 4136-4160.
[2] Gong, Zixuan, et al. "NeuroClips: Towards High-fidelity and Smooth fMRI-to-Video Reconstruction." NeurIPS 2024.