Skip to content

Latest commit

 

History

History
27 lines (22 loc) · 1.9 KB

README.md

File metadata and controls

27 lines (22 loc) · 1.9 KB

Training Datasets

We provide the Colmap 3D reconstruction results and the filtered camera movement sequences in our DroneMotion-99k dataset. You can download either a minimal dataset with 10 videos and 129 sequences OneDrive link or the full dataset with 13,653 videos and 99,003 camera trajectories OneDrive link. We also provide them in the Huggingface Dataset.

After downloading the training data, your folder should look like this

dvgformer/
├── youtube_drone_videos/
│   ├── dataset_full.h5
│   └── dataset_mini.h5
├── src/
├── README.md
...

Due to the YouTube policy, we cannot share the video MP4s or the frames. As an alternative, we include a python script download_videos.py that can help you automatically download the videos and extract the frames.

python download_videos.py --hdf5_fpath youtube_drone_videos/dataset_mini.h5
python download_videos.py --hdf5_fpath youtube_drone_videos/dataset_full.h5

This should update your downloaded HDF5 dataset file with the video frames.

You can also adjust the number of workers for the download process or the frame extraction process in download_videos.py by specifying --num_download_workers or --num_extract_workers.

Tips

  • Please keep a stable network connection during the download. Also avoid termination of the python script download_videos.py, which can corrupt the updated HDF5 file as in issue #1.
  • Following the above, you might want to backup the dataset_full.h5 file prior to running the download script, as the corruption is irreversible and might require a re-download of the original HDF5 file.