about how to use cv2:: codacodec:: VdieoReader::set()

[RookieDeng33]

How to use similar functions in cv2:: codacodec:: VdieoReader like cv2:: VideoCapture:: set (CAP_PROP_POS_FRAMES, target)

I need to read the video frame from the customized key frame position, or repeat the reading from the specified frame when reading the video to the last frame

just like

def get_cap_frames_by_range(f_cap, frames_idx, prepared_f_frames):
      min_index, max_index = min(frames_idx), max(frames_idx)
      # print(min_index, max_index)
      temp_counter = {}
      temp_frames = {}  
      for item in frames_idx:
          if item not in temp_counter:
          temp_counter[item] = 0
          temp_counter[item] += 1
      if int(f_cap.get(cv2.CAP_PROP_POS_FRAMES)) != min_index:
          # print("switch CAP_PROP_POS_FRAMES")
          f_cap.set(cv2.CAP_PROP_POS_FRAMES, min_index)
      ...

[asmorkalov]

cc @cudawarped

[cudawarped]

@RookieDeng33 You can seek to an initial frame position when initializing the decoder, see this test, however
for efficient decoding cudacodec::VideoReader uses a separate thread to demux the video source and queue the frames for decoding. This means that it is not possible to seek to a specific frame during decoding process.

If there is a small gap between the frames you require then I would just decode and skip the intermediate frames. This could be equivalent to seeking if the gap between those frames is less than the gap between key frames. e.g. If you want frames 10,20,30,...,100 but the key frame is every 250 frames the internal logic when seeking would have to decode all of the frames from frame 0 anyway making the current behaviour of cudacodec::VideoReader equivalent to cv::VideoCapture.

[RookieDeng33]

@RookieDeng33 You can seek to an initial frame position when initializing the decoder, see this test, however for efficient decoding cudacodec::VideoReader uses a separate thread to demux the video source and queue the frames for decoding. This means that it is not possible to seek to a specific frame during decoding process.

If there is a small gap between the frames you require then I would just decode and skip the intermediate frames. This could be equivalent to seeking if the gap between those frames is less than the gap between key frames. e.g. If you want frames 10,20,30,...,100 but the key frame is every 250 frames the internal logic when seeking would have to decode all of the frames from frame 0 anyway making the current behaviour of cudacodec::VideoReader equivalent to cv::VideoCapture.

Thank you for your reply, which has helped me a lot; However, I still have some doubts. Why not reset the queue but recreate the cudacodec:: VideoReader instance; For video files, is it more efficient to reset the queue than to reload the video source?

[cudawarped]

Why not reset the queue but recreate the cudacodec:: VideoReader instance; For video files, is it more efficient to reset the queue than to reload the video source?

I am not clear on your what you are asking?

To clarify I am not suggesting you create a new instance of cudacodec::VideoReader for each frame when you want to seek. I am informing you that you cannot seek with cudacodec::VideoReader due to the way it is currently implemented, but you can if you want choose the index of the first frame to start your decoding session from.

I am also pointing out that when you skip to a frame in between key frame the decoder has to decode every frame from that key frame to your desired frame anyway. This means that depending on your use case (how close together the frames you want to skip to are and the GOP size etc.) you can simply decode all the frames and ignore the ones you don't need.

[RookieDeng33]

Why not reset the queue but recreate the cudacodec:: VideoReader instance; For video files, is it more efficient to reset the queue than to reload the video source?

I am not clear on your what you are asking?

To clarify I am not suggesting you create a new instance of cudacodec::VideoReader for each frame when you want to seek. I am informing you that you cannot seek with cudacodec::VideoReader due to the way it is currently implemented, but you can if you want choose the index of the first frame to start your decoding session from.

I am also pointing out that when you skip to a frame in between key frame the decoder has to decode every frame from that key frame to your desired frame anyway. This means that depending on your use case (how close together the frames you want to skip to are and the GOP size etc.) you can simply decode all the frames and ignore the ones you don't need.

Sorry for not describing my problem clearly. Parameter frames_idx of function get_cap_frames_by_range is a batch of consecutive video frame idx. For example, the first batch of frames_idx is [1,2,3,4,5], and the second batch may be [4,5,6,7,8] , and the third batch may be [9,10,11,12]. I mean that when decoding each batch, it may need to create a new instance of cudacodec::VideoReader, but not necessarily.
What I asked earlier is that when reading from the specified firstFrameIdx, you may use the instantiated VideoReader to clear The FrameQueue，and then re put the frames which start with firstFrameIdx to FrameQueue for decoding.is it more efficient than create a new instance of cudacodec::VideoReader?

[cudawarped]

@RookieDeng33 I still don't fully understand what you are asking. That said if you altered the implementation so that you could seek to a specifiec location within the file during decode then it should be more efficient to do that than create a new instance of cudacodec::Videoreader.