Object Detection

The goal of this project is to recognize and to localize objects of a given category in arbitrary still images. We particularly focus on the problem of within-class object variation exemplified on the right. The method combines machine learning techniques and local image features (not "interest points").

This work is motivated by remarkable success of many method using histogram based image representations: [Swain & Ballard 1991] [Schiele & Crowley 1996] [Lowe 1999] [Schneiderman & Kanade 2000] [Leung & Malik 2001] [Belongie et.al. 2002] [Dalal & Triggs 2005]

We focus on selecting support regions of histogram image features for a given object class. As illustrated on the left, information within different object regions may vary in terms of discriminative power (regions "A" are more discriminative than regions "D"). Similar to [Levi & Weiss 2004] we use AdaBoost to select histogram regions optimized for the classification of training samples.

We compute histograms of image gradient discretized in four orientations. During training the histograms are computed for rectangular object regions with all possible positions, sizes and shapes. Fisher weak learner is then applied to select a histogram feature and an associated classifier at each round of AdaBoost.

We found particularly useful to increase the training set by applying random small affine transformations to training images. This step makes the detection process more robust to corresponding transformations in test images.

We detect objects by classifying densely sampled sub-windows (with fixed aspect ratio) of a given test image ("sliding window" method). Multiple detections with similar positions and sizes are grouped by clustering. The size of a cluster is considered as the detection confidence.

Resources

Paper

- Improvements of Object Detection Using Boosted Histograms (2006),
I. Laptev; in Proc. BMVC'06, Edinburgh, UK, pp. III:949-958.

Slides

PowerPoint talk at BMVC 2006, Edinburgh

© 2008 TT-solutions

• About TT-Solutions