-
Python libraries like opencv, and matplotlib should be installed
-
Install cocoapi (make sure u can run coco's demo code)
-
FFmpeg is also needed for combining videos
-
Separate each video into frames (with ffmpeg); frames from the same video need be in the same folder and not mixed up with another other images or frames from other videos & the names of the folders should be the same as the video. The frames themselves should also be stored as
<video_name>-<frame 3>.jpg- I store all frames in the folder
shark_frames, and inshark_frames, there is a folder for each video w/ the same name as video. (eg. frames fromDJI_0386.mp4are stored inshark_frames/DJI_0386/DJI_0386-<frame #>.jpg)
- I store all frames in the folder
-
VIA annotations should have the same name as their video counterparts; via's meta data isn't considered.
- The annotations should have object label or object present as well as an id. If another word is used, change the code around here
- VIA to Coco: Here's a sample notebook for converting a single video's VIA to Coco.
- Alternatively, you can use the commandline version to convert a whole folder of annotations; type
python via2coco.py --helpfor more info.
- Alternatively, you can use the commandline version to convert a whole folder of annotations; type
- Coco to py-MDNet input: This converts coco annotations to py-MDNet's input json format; it will generate a file for each object
create_bash.pywas used to generate bash scripts to submit usingsbatch- This notebook draws the bounding boxes using the input and output files from py-MDNet.