Object Detection Under Clutter With Visual Modeling Approach

Object detection under clutter is a challenging task of vision research. Illumination and observer varaiations result in the complex variations of background and object. Imaging from 3D to 2D makes this worse because of noise incurring and information loss.In this paper, we systematically explore the task of object detection under clutter, involving with: the methodology base of vision research, 2D object detection on single frame(including Bag of Visual Words Model, Pictorial Model), 2D object detection on sequence(Na?ve Bayesian Model). We not only emphasize thinking on theory and approaches, but also emphasize practising on algorithms. Our research work is briefly depicted below.1. We conclude the visual modeling approach based on Marr’s representation theory of vision and statistical learning theory. This approach has two important properties: emphasizing research on both computations and physical constraints; emphasizing vision is essentially a probabilistic inferring process. This approach is the methodology base of this paper, which fundamentally guides our research.2. We analyze cortex-like HMAX model from both computational theory and algorithmic designing, and point out that HMAX is a Bag of Visual Words Model in essence. Based on this analysis, we improve HMAX in two sides: introducing bottom-up attention constraint into the process of visual words selection; proposing discriminative shape model through combining HMAX and explicit shape model. In fact, an extra constraint – object shape – is introduced by discriminative shape model. Experiments show the effectiveness of our improvements.3. We analyze Pictorial Model with visual modeling approach. For one kind of design of Pictorial Model – star shape DPM, we propose a new cortex-like feature – HOGabor based on Gabor atomic primitive. Experiments on PASCAL VOC 2007 etc show that HOGabor can significantly improve the performance on rigid object like aeroplane and car etc. We comprehensively evaluate DPM on multiple datasets and analyze the relationship between sample, model complexity and model generalization.4. For another kind of design of Pictorial Model – tree shape POSE, we point out its three shortcomings: only considering the part marginal; error for scale estimation; computational complexity. We propose two schemes to deal with these problems: obtaining maximal posterior(MAP) of the object; searching the object based on root part. Experiments show the performance and efficiency of human pose estimation can be improved by these two schemes.5. For continuous visual input, we set up a practical Bauesian model based on offline-learned object prior and online-learned likelihood of biased saliency. We adopt the conditional independence assumption of features, and can use the same low level feature-- Gabor atomic primitive. In this model, we demonstrate the interactions between bottom-up and top-down procedures, also we reflect the coarse-to-fine procedure of object detection.In this paper, we firstly conclude the the methodology base of vision research--visual modeling approach; then systematically explore three classes of 2D object detection model under the guide of visual modeling approach. We improve or recreate these models, some of our work is ready for industrial use, and some has academic value. We have obtained coarse-to-fine representations for 2D object, which are foundations for approaching 3D representations for object.