Documentation

Setup

Python 3 is assumed to be the version of python in use.

To install the required packages run:

pip install -r requirements.txt

Add ffmpeg with

sudo apt-get install ffmpeg

Video download & Crop

Video Download

To download a video from youtube you have to:

• Create a videos folder
• Create a links.txt file containing the videos you want to download
• Run python download_video.py

The videos will be downloaded into the videos directory.

Video Crop

To crop an already downloaded video you can run:

python crop_video.py {VIDEO_TO_CROP} \
    -o {CROPPED_VIDEO} \
    --start {START_TIME_IN_SECONDS} \
    --end 20 {END_TIME_IN_SECONDS}

Example - Crop video and retain the part between 5-15 seconds, saving its output to mydata/cropped.mp4:

python crop_video.py videos/full_video.mkv \
    -o videos/cropped.mkv \
    --start 5 \
    --end 15

Frames extraction

To extract frames from an already downloaded video you can run

    python extract_frames.py {PATH_TO_VIDEO} --start {START_TIME in format hh:mm:ss} --end {END_TIME in format hh:mm:ss}

Example - Extract frames from 1'25" to 1'49"

python extract_frames.py videos/76ers_vs_nuggets_dec2019.mp4 --start 00:01:25 --end 00:01:49

Highlight extraction with OCR

To extract highlights from an already downloaded video you can run

    python ocr.py --input {PATH_TO_VIDEO} --ocr True

Example - Extract highlights from video raptors_warriors_2019.mp4

python ocr.py --input raptors_warriors_2019.mp4 --ocr True

The aforementioned video can be downloaded from here: https://drive.google.com/open?id=1jEzUPWSsGkL9jn4K0osA5dlDunhr_Cuj

Annotation format transformation

You can transform YOLO to VOC and vice versa using the format_transform/format_transform.py script.

Granted that you have annotated your images inYOLO format and saved the images and their annotations in a {DATASET} directory, you can add VOC format annotations by executing:

python format_transform.py {PATH_TO_DATASET_DIRECTORY} yolo_to_voc

If you want to do the VOC to YOLO transformation you can execute:

python format_transform.py {PATH_TO_DATASET_DIRECTORY} voc_to_yolo

The new annotations will be saved along with the images and the initial anotations at the {DATASET} directory.

Note that, along with the images and the annotations, a classes.txt file must be present at the {DATASET} directory. Using labelimg to annotate the images will, normally, lead to automatic creation of this file.