A Bayesian framework for object recognition under severe occlusion

Abstract

Shape classification has multiple applications. In real scenes, shapes may contain severe occlusions, hardening the identification of objects. In this work, a bayesian framework for object recognition under severe and varied conditions of occlusion is proposed. The proposed framework is capable of performing three main steps in object recognition: representation of parts, retrieval of the most probable objects and hypotheses validation for final object identification. Occlusion is dealt with separating shapes into parts through high curvature points, then tangent angle signature is found for each part and continuous wavelet transform is calculated for each signature in order to reduce noise. Next, the best matching object is retrieved for each part using Pearson’s correlation coefficient as query prior, indicating the similarity between the part representation and of the most probable object in the database. For each probable class, an ensemble of Hidden Markov Model (HMM) is created through training with the one-class approach. A sort of search space retrieval is created using class posterior probability given by the ensemble. For occlusion likelihood, an area term that measure visual consistency between retrieved object and occlusion is proposed. For hypotheses validation, a area constraint is set to enhance recognition performance eliminating duplicated hypotheses. Experiments were carried out employing several real world images and synthetical generated occluded objects datasets using shapes of CMU_KO and MPEG-7 databases. The MPEG-7 dataset contains 1500 test shape instances with different scenarios of object occlusion with varied levels of object occlusion, different number of object classes in the problem, and different number of objects in the occlusion. For real images experimentation the CMU_KO challenge set contains 8 single view object classes with 100 occluded objects per class for testing and 1 non occluded object per class for training. Results showed the method not only was capable of identifying highly occluded shapes (60%-80% overlapping) but also present several advantages over previous methods. The minimum F-Measure obtained in MPEG-7 experiments was 0.67, 0.93 and 0.92, respectively and minimum AUROC of 0.87 for recognition in CMU_KO dataset, a very promising result due to complexity of the problem. Different amount of noise and varied amount of search space retrieval visited were also tested to measure framework robustness. Results provided an insight on capabilities and limitations of the method, demonstrating the use of HMMs for sorting search space retrieval improved efficiency over typical unsorted version. Also, wavelet filtering consistently outperformed the unfiltered and sampling noise reduction versions under high amount of noise.