Option to generate video JSON with all frames instead of only the best of that category

[PetervanLunteren]

For the EcoAssist and Timelapse integration of video processing results, we need an option to get a video JSON file with detection results for all frames processed.

Now I do it inside EcoAssist like so:

    # Load the video recognition JSON file
    if video_json:
        with open(video_json, 'r') as video_file:
            video_data = json.load(video_file)

        # group the frame-level detections into a their parent video
        if re.search(r"frame\d{6}\.jpg", video_data['images'][0]['file']):
            aggregated_detections_per_video = defaultdict(list)
            for frame in video_data['images']:
                video = os.path.join(*Path(frame['file']).parts[:-1])
                frame_number = int(Path(frame['file']).stem[5:])
                detections = frame['detections']
                detections = detections[0]
                detections['frame_number'] = frame_number
                aggregated_detections_per_video[video].append(detections)

            new_video_data = []
            for video, detections in aggregated_detections_per_video.items():
                new_video_data.append({"file": video, "detections": detections})

            video_data['images'] = new_video_data

That would result in something like this:

  "file": "vid1.mp4",
   "detections": [
    {
     "category": "1",
     "conf": 0.944,
     "bbox": [
      0.5789,
      0.5,
      0.3414,
      0.3249
     ],
     "classifications": [
      [
       "28",
       0.98437
      ]
     ],
     "frame_number": 0
    },
    {
     "category": "1",
     "conf": 0.94,
     "bbox": [
      0.6335,
      0.5152,
      0.2757,
      0.3111
     ],
     "classifications": [
      [
       "28",
       0.98423
      ]
     ],
     "frame_number": 30
    },
    ...

But it feels quite cumbersome to convert this inside EcoAssist (and error prone), if MegaDetector had removed them earlier. What are you're thoughts about this?

[agentmorris]

@PetervanLunteren, does the --include_all_processed_frames option to process_video.py already do what you're looking for?

If you already have a .json file that came from somewhere else, and you want to use the same mechanics, you can use the frame_results_to_video_results function, which is what process_video uses to convert frame-level results to video-level results. It takes an options object that has an include_all_processed_frames option.

Let me know if that already does what you're looking for, or if not, how the behavior you're looking for is different?

Either way, I think we're between 90% and 100% of the way there.