Factorization for Probabilistic Local Appearance Models

B. Moghaddam and X. Zhou (USA)



We propose a novel local appearance modeling method for object detection and recognition in cluttered scenes. The approach is based on the joint distribution of local feature vectors at multiple salient points and factorization with Independent Component Analysis (ICA). The resulting non-parametric densities are simple multiplicative histograms. This leads to computationally tractable joint probability densities which can model high-order dependencies. Furthermore, we propose a distance-sensitive histograming technique for capturing spatial dependencies which are otherwise lost in the joint feature distributions. The advantages over existing techniques include the ability to model non-rigid objects and the flexibility in modeling spatial or structural relationships between object parts. Testing and evaluation shows that the factorized density model with spatial encoding improves modeling accuracy and outperforms global appearance models in image/object retrieval. Furthermore, experiments in detection of substantially occluded objects in cluttered scenes have demonstrated promising results.

Important Links:

Go Back