Taxonomic classification for web-based videos

Taxonomic classification for web-based videos

Author: Yang Song et al. (Google)Presenters:

Phuc Bui & Rahul Dhamecha

1. Introduction

Taxonomic classification for web-based videos

Web-based Video Classification

• Web-based Video (e.g. Youtube)– Over 800 million unique users visit / month– Over 4 billion hours of video are watched / month– 72 hours of video are uploaded / minute

• Classification– Improve User experience– Increase Website profit

What’s interesting?

• Large-scale classification– Taxonomy of categories– Unlimited domain

• Combined Approach– Text

• Labeled Web documents• Labeled Video

– Video• Content-based features

Overview Approach

• Multi-labels Classification– One classifier for each category

• Classifiers– Text-based Classifier• from Web-based Documents

– Combined Classifier• Text-based Classifier• Video content-based features

2. Algorithms

TAXONOMIC CLASSIFICATION: - THE VARIOUS CATEGORIES.

TRAINING SET OF EACH CATEGORY

Pre-trained text based classifiers of each category used for porting videos Labeled Video data is used for training these classifiersNo. of Classifiers = No. of CategoriesAda-boosting is deployed to aggregate these weak classifiers to a Strong Classifier

MIGRATION FROM TEXT TO VIDEO

Feature Extraction

Text Based Features.

President Obama: the Real Mitt Romney - Denver, Colorado

Title

Description

Keywords

Content Based Features

Moments from multi-scale analysis

Color HistogramMean, variance of each channel.

Difference between mean of center and boundary

Content Based Features contd…

Edge DetectionCanny Edge Detection Algorithm

Content Based Features contd…

Color Motion Features

Cosine Difference of the histograms of subsequent frames.

Content Based Features contd…

Shot Boundary Features

TypesHard CutFadeDissolveWipe

Hard Cut

instantaneous transition from one scene to the next

Fade

A Fade which is a gradual between a scene and a constant image (fade-out) or between a constant image and a scene (fade-in).

Dissolve

A Dissolve is a gradual transition from one scene to another in which the first scene fade-out and the second scene fade-in. so it is a combination of fade-in and fade-out.

Wipe

A Wipe is a gradual transition in which a line move across the screen, with the new scene appearing behind the line.

Integration

Labled Videos

Text Based Feature

Extraction

Apply Pre-trained Text Classifiers

F Score from Classifiers

Labled Videos

Content Based Feature

Extraction

F Score and Content Based Features are

combined

A new Classifier is trained.

3. Experiments

Data

• 5789 videos• 9087 labels• 565 categories

• 80% training• 20% evaluation

Evaluation

• Precision

• Recall

• F-score

Results

• Sample videos

Results

• 80-category classifiers • 1037-category classifiers

Results

Results

• Adaption + Content-based features classifiers

• Content-based features-only classifiers

4. Conclusion & Dicussion

• Video features– Content-based– Associated texts

• Web-documents based text classifier

• Semi-supervised learning

• Image-based classifiers– ImageNet