Home
TV stream classification using visual information involves analyzing and categorizing television content based on what is currently being displayed on the screen. Media agencies use this approach to audit TV streams in order to make informed decisions about advertising, programming, and audience engagement. By using visual information to classify TV streams into following three categories, media agencies can effectively audit and analyze content for advertising opportunities, audience preferences, and programming strategies:
This category focuses on identifying commercial segments within the TV stream. Visual cues, such as the presence of a product or brand logos, promotional text, specific editing styles, or recognizable jingles, can help differentiate advertisements from other content types. Advertisements are typically short in duration and designed to grab viewers' attention, making them visually distinct from news or regular programs.
News content is characterized by its informative and factual nature, often covering current events, politics, weather, sports, or human-interest stories. Visually, news segments can be identified by the presence of on-screen elements, such as newsroom backgrounds, news tickers, lower-thirds displaying headlines or names of reporters, and video footage or images relevant to the story being covered. News anchors, reporters, and interviewees are often featured in these segments, adding another layer of visual identification.
This category encompasses all non-news and non-advertisement content, including TV shows, movies, documentaries, and other forms of entertainment. Programs can be visually classified based on factors such as the presence of recurring characters, distinctive settings, and specific storytelling or visual styles. Additionally, on-screen graphics, such as opening credits, show titles, or episode numbers, can aid in identifying and categorizing the content.
The subject repository is responsible for Stream Classification. Final classification is concluded on the basis of temporal probabilities & implemented as server-client architecture. Currently three classes can be classified using subject architecture.
- Advertisement
- News
- Program
The model used in this architecture is ConvNext-Tiny, finetuned using Pytorch framework for Pakistani news channels.
