Analysis-Of-News-Broadcast-Video
Computer Vision Project
Authors: Hantang Li, Hongyi Sun
Analyzed News Broadcast Videos and Implemented shot detection, Logo Detection, Face Detection and Face Association algorithms that can be used in news video editing software. With some hypotheses and experiments on ways to improve some algorithms. I built a shot detection algorithm by evaluating the local variance of histogram difference, the algorithm performed better on news clips than thresholding colour histogram and edge change ratio. I also implemented two face association algorithms based on the cost of the face
To produce the final result, please follow the following steps:
- Run notebook file [detection.ipynb]
- At the end, download the .zip file generated by the code The .zip file contains frames with detected faces/gender Now we use those frames to detect logo.
- For clip 1 run [logo_template_match.py] only, and use it to detect "logo2.png".
- For clip 2, run [logo_template_match.py] first to detect "logo_clevver.jpg", then run [logo_feature_match.py] to detect "nbc_clip2.jpg"
- For clip 3, run [logo_template_match.py] only, and use it to detect "logo_flick.jpg".
detection.ipynb: This file contains codes for part (f), (g) and (i). By running all the code cells (except for the ones I noted \not to run), you will generate a .zip file for processed frames for Clip 2.
If you want to detect other clip, simply change all occurence in the code of "clip 2" to "clip x". Where x in [1, 2, 3]
train_face_detection.ipynb: This file contains codes to train the face detector using Yolov5
- CelebA dataset. Note: Downloading CelebA dataset and extracting can take up to 20 hrs. Training Yolov5 medium on 70k dataset size takes roughly 50 min per epoch on p100 GPU.
logo_template_match/logo_feature_match.py: These 2 files contains codes to detect logo using template and feature matching. Please change the corresponding logo & image path to your local path.