sebothetramp/open_video_helmet

Open Video Molina

This is a growing open-source repository for multiview video captured across different places. More locations and recording sequences may be added over time.

The current release contains GoPro recordings from three locations: forest, predaia, and molina. Each location has six camera views (camera_00 through camera_05). The GoPros are mounted together on a fixed multicamera rig, keeping their relative arrangement approximately consistent during each capture.

The videos are encoded as HEVC at 2704x2028, 30000/1001 fps, with stereo AAC audio at 48 kHz.

Data

The uploadable data is under videos/:

videos/
├── metadata.csv
├── forest/
│   └── forest_camera_00.mp4 ... forest_camera_05.mp4
├── predaia/
│   └── predaia_camera_00.mp4 ... predaia_camera_05.mp4
└── molina/
    └── molina_camera_00.mp4 ... molina_camera_05.mp4

The local raw/ directory contains coherently renamed safety copies of all 36 source clips. It is excluded from upload by .gitignore to avoid duplicating approximately 93 GB.

Temporal structure

• —forest: two contiguous GoPro chapters merged without re-encoding.
• —predaia: one source clip, copied without re-encoding.
• —molina: two contiguous GoPro chapters followed by a later recording from the same location, merged without re-encoding. The inactive interval before the third component is not represented in the merged timeline.

The views have matching recording durations, but the cameras' embedded clocks are not perfectly synchronized. No frame-level synchronization or temporal alignment has been applied.

Loading

from datasets import load_dataset

dataset = load_dataset("videofolder", data_dir="videos", split="train")

Upload

After creating an empty Hugging Face dataset repository:

HF_XET_HIGH_PERFORMANCE=1 hf upload USER/REPOSITORY /data0/sebastian.cavada/open_video_molina . --repo-type=dataset

See SOURCE_MAPPING.md for the exact source-to-output mapping.