Leveraging Artificial Intelligence for Scene Detection in Broadcast Video
summary
Claims a fully automated scene detection pipeline that segments videos into story units without a human editor and places ad breaks. Implementation trains a deep network to learn a distance metric over shots using 2048-dim Inception bottleneck visual descriptors, mel spectrogram audio encodings and Google News word2vec text embeddings, produces a cosine similarity matrix over all shots, clusters contiguous shots and picks local minima as boundaries with confidence scores. Trained on BBC Planet Earth, it shows clear diagonal clusters and tunable precision for obvious versus subtle scene breaks.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:08Why Scene Detection
- 2:27Architecture
- 4:16Neural Network
- 6:44Textual Network
- 8:12Single Feature Factor
- 9:35Clustering Strategy
- 11:05Scene Breaks
- 12:44Confidence Scores
related talks
42mDeep Learning for Automatic Basketball Video Production
35mFootball video analysis using Deep Learning
27mWhen the Art of Entertainment ties the knot with Science
37mHow we generate 130,000 shorts per month using AI
35mLarge-Scale Video Intelligence
28mScalable High Precision OCRs for Game Event Detections
40mSmart Crop and Smart Preview via Video Understanding
17mAI/ML demo derby