Object Recognition with 3D Models

Intro: We propose several new ideas of how to use 3D models for viewbased object recognition. In an initial experiment we show that even the simple task of distinguishing between two objects requires large training sets if high accuracy and pose invariance are to be achieved. Using synthetic image data, we propose a method for quantifying the degree of difficulty of detecting objects across views and a novel alignment algorithm for pose-based clustering on the view sphere. Finally, we introduce an active learning algorithm that searches for local minima of a classifier’s output in a low-dimensional space of rendering parameters. Experimental setup: Synthetic training and test images were rendered from five textureless 3D models (see fig. 1) by moving a virtual camera on a sphere around each model. The models were illuminated by ambient light and a point light source. The six free rendering parameters were the camera’s location in azimuth and elevation, its rotation around its optical axis, the location of the point light source in azimuth and elevation, and the intensity ratio between ambient light and the point light. The rendered images were converted to 23×23 grayvalue images. From those we extracted 640 dimensional vectors of histograms of gradients. Our main classifier was an SVM with a Gaussian kernel.