extract frames from mp4 pipeline

[Joe-Surrey]

I wanted to extract frames from an mp4 file so I wrote this
Sharing in case anyone finds it useful (or knows a better way)

@PIPELINES.register_module()
class LoadImageFromMP4:
     """Loading MP4 from file.
    image_file in json must be in the format mp4file%frame
        e.g. car_video.mp4%6 for frame 6 of car_video.mp4
    """

    def __init__(self,):
        self.readers = {}


    def __call__(self, results):
        """Loading image from file."""
        video_file,frame = results['image_file'].split("%")
        if video_file not in self.readers:
            self.readers[video_file] = mmcv.VideoReader(video_file)
        
        img = self.readers[video_file].get_frame(int(frame))

        if img is None:
            raise ValueError('Fail to read frame {} of {}'.format(frame,image_file))

        results['img'] = img
        return results

[innerlee]

If the dataset is very large, retaining so many reader objects might use out the OS's number of opened files quota, right?

[Joe-Surrey]

yeah maybe it would be good to delete the reader after all of its frames have been read