Research On Multi-object Detection Algorithms Utilizing Local Features And Spatial Context Cues

Recent breakthroughs in Communicational Technology, especiallyNetwork Technology, together with the widespread applications ofintelligent systems contribute to the dramatic changes of people s preferencein communication over Internet from plain text to images and videos.Therefore, it is necessary to investigate the rich content of those data toobserve the latest trends, to understand the hidden meanings and to benefitcomprehensive decision-making. Meanwhile, the reliable detection result ofall the objects in the image or video offers a solid background for furtherunderstanding the semantics of images and videos. In addition, object resultalso benefits in solving other problems in the field of computer vision, suchas semantic segmentation and object parsing. However, currentstate-of-the-art object detection system still cannot offer a satisfyingperformance in real-world images as the objects varies in size, pose andsuffers from viewpoint variation, uneven illumination, truncation andocclusion. This paper researches the object detection algorithms utilizinglocal features and spatial context cues.This paper first analyzes the object detection model utilizing local cuesand the model combining local and context cues. By implementing thestructural model regarding local features and spatial context information,this paper further investigates the merits and drawbacks of current structuralmodel. Due to the fact that the number of object categories might be huge while the relationship between objects can be extremely varied, the trainingand inference stage of structural model is computational intensive. Thus thispaper proposes isolation methods to separate the structural model into a setof binary classification problems, including threshold method and maximalmethod. Experiment shows that isolation method could have huge speedgain at the cost of little performance loss. Meanwhile, the model couldcompensate the performance loss via taking more spatial contextinformation, e.g. relative aspect ratio, overlap percentage, etc.It is common for current object detection method with local and contextcues to use fixed, pre-defined spatial relationships. However, this kind ofspatial relationship might not be accurate or semantic. This paper proposes adata-driven method to discover the spatial context information for all objectpairs. This method adopt clustering algorithm on the relative displacementbetween objects to form spatial context prototypes (SCPs), and build thespatial context features by SCP. Finally, a structural model is used toimprove the detection result. Instead of K-Means algorithm, this paperproposes contrast clustering algorithm (CKM) to discover the SCP.Experiment on PASCAL VOC2007dataset also manifest that CKM worksbetter than K-Means algorithm. This paper also examines different codingmethods in constructing spatial context features. The coding method withoptimal performance in validation set is chosen to construct the spatialcontext feature and improve the model accuracy.After obtaining SCP, this paper also optimizes the utilization of spatialcontext information. Based on the observation that the spatial contextpatterns differ between different classes, it also differs between differentposes for same object pairs. This paper proposes a spindle model todistinguish different poses of a same object and further improves the modelperformance.