SINETRA

Helpers to handle Sinetra dataset [11].

We only provide functions to load annotations.

byotrack.dataset.sinetra.load_metadata(path: str | os.PathLike) → dict[str, torch.Tensor]

Load all the metadata of the ground-truths.

It loads the outputted metadata from SINETRA (position, size, rotation and intensity weight) See: https://github.com/raphaelreme/SINETRA

JSON Format:

{
    "mu": torch.Tensor (T, N, D)  # Position
    "std": torch.Tensor (T, N, D)  # Size
    "theta": torch.Tensor (T, N, 1 or D)  # Rotation (1 in 2D, 3 in 3D)
    "weight": torch.Tensor (T, N)  # Weight
}

Parameters:

path (str | os.PathLike) – Path to the generated folder or to the video_data.pt file.

Returns:

Metadata (position, size, rotation, and intensity weight)

Return type:

dict[str, torch.Tensor]

byotrack.dataset.sinetra.load_tracks(path: Path) → list[Track]

Load ground-truth tracks, built from the metadata.

This is quite simple, it uses only the positional metadata (“mu”) to build tracks. Each track is defined from frame 0 to the end in current implementation of SINETRA.

Parameters:

path (str | os.PathLike) – Path to the generated folder or to the video_data.pt file.

Returns:

Ground-truth tracks

Return type:

list[byotrack.Track]