How to use Python decord library to seek by timestamp instead of by frame index?

Question

Decord allows to seek a video frame from a file using indices, like:

video_reader = decord.VideoReader(video_path)
frames = video_reader.get_batch(indices)

How can I do the same if I have timestamps (with the unit second)?

Answer 1

You can obtain the timestamp of each frame (averaging its start and end times), then find the closest one:

from typing import Sequence, Union

import decord
import numpy as np


def time_to_indices(video_reader: decord.VideoReader, time: Union[float, Sequence[float]]) -> np.ndarray:
    times = video_reader.get_frame_timestamp(range(len(video_reader))).mean(-1)
    indices = np.searchsorted(times, time)
    # Use `np.bitwise_or` so it works both with scalars and numpy arrays.
    return np.where(np.bitwise_or(indices == 0, times[indices] - time <= time - times[indices - 1]), indices,
                    indices - 1)

frames = video_reader.get_batch(time_to_indices(indices))

Note the VideoReader C object (not the Python one) is loaded already with all the frame timestamps from the initialization. And we take advantage of the fact that they should be sorted.