ramyamut/dhcp_fmri_fetal_pose

Here we provide the manual annotations (brain/eyes segmentations and poses) used to train and evaluate methods for head pose estimation from the publicly available dHCP¹ fetal brain BOLD fMRI volumes. Specifically, we annotated the first frame in each time-series, resampled to 3mm isotropic voxels, for each subject.

The data is organized in the following structure:

├── training
│   ├── images
│   │   ├── subjectXXX.nii.gz
│   │   └── ...
│   ├── masks
│   │   ├── subjectXXX.nii.gz
│   │   └── ...
├── validation
│   ├── images
│   ├── masks
├── testing
│   ├── images
│   ├── masks
├── annotations.csv

Manual segmentations can be found in the masks/ subdirectories. Labels 1 and 2 correspond to the brain and eyes, respectively.

Manually labeled poses can be found in the file annotations.csv. It has the following format:

    | frame_id | fold | rot_x | rot_y | rot_z |
    |----------|------|-------|-------|-------|
    | ...      | ...  | ...   | ...   | ...   |

where frame_id is the file name of the volume without the file extension; fold is either "training", "validation", or "testing"; and rot_x, rot_y, rot_z are the Euler angles in degrees of the rotation from the volume to the canonical coordinate frame. The Euler angle rotation assumes the "xyz" ordering convention.

¹ The developing human connectome project fetal functional MRI release: methods and data structures
Karolis et al
Imaging Neuroscience, 2025