Generative part-based Gabor object detector

• Probabilistic local part descriptor using complex-valued Gabor feature bank.
• Alternative approach using a large Gabor filter bank and part specific optimization.
• Randomized Gaussian mixture model robust to a small number of training examples.
• Probabilistic part constellation model in “aligned object space”.
• Part-based “pose quantization” that provides pose invariant detection.

Discriminative part-based models have become the approach for visual object detection. The models learn from a large number of positive and negative examples with annotated class labels and location (bounding box). In contrast, we propose a part-based generative model that learns from a small number of positive examples. This is achieved by utilizing “privileged information”, sparse class-specific landmarks with semantic meaning. Our method uses bio-inspired complex-valued Gabor features to describe local parts. Gabor features are transformed to part probabilities by unsupervised Gaussian Mixture Model (GMM). GMM estimation is robustified for a small amount of data by a randomization procedure inspired by random forests. The GMM framework is also used to construct a probabilistic spatial model of part configurations. Our detector is invariant to translation, rotation and scaling. On part level invariance is achieved by pose quantization which is more efficient than previously proposed feature transformations. In the spatial model, invariance is achieved by mapping parts to an “aligned object space”. Using a small number of positive examples our generative method performs comparably to the state-of-the-art discriminative method.

Figure optionsDownload high-quality image (295 K)Download as PowerPoint slide