| As a quite challenging problem, object detection is difficult because of the divergences among objects of different categories. The traditional Deformable Parts Model in a way solved the view changes and deformation problem by training components of different shooting views and using a set of deformation costs to connect each part filter to the root filter. However, the model is short of the use of the dependencies among categories and the hidden location of objects, thus the improvements of the performance gradually run into the bottleneck. This paper analyzes different methods to improve the performance of DPM throug h feature extracting, filters setting of different parts and optimization methods, also introduces a cascaded training process to integrate the information above, therefore the confidence scores of the detection results are re-evaluated:First, this paper analyzes several kinds of features extracted from the dataset. During the DPM detecting procedure, low-dimension Hog, Sparse Codes and CNN features are compared for their expression power and computing speed s. In the rescore procedure, features such as Gist, Geometric and CNN are used to provide reliable sources of information for the following steps.Secondly, the detecting process of DPM is explored, as the root and parts filters are matched, and the locations of parts are set through a dynamic programming progress. The final cost function includes the matching scores of filters and the deformation cost of parts and constants. Issues as the numbers of parts, the set of different component and the solutions of the cost function are emphatically discussed.Finally, based on the detection results of DPM, different training targets are set according to different forms of the potential information, like categories, locations and co-occurrence, then the method uses these targets and the features above to train SVM classifiers. Combining these results with the detection results of DPM and spatial limits, the method can train a re-score model which produces a new detection score for each detection results, through which the potential information left impacts on the detection results, and better detection results are obtained.The method is tested on PASCAL VOC 2007 and INRIA datasets, experimental results show that DPM method could be further improved, and using potential information from specified targets to re-score the detection results could somehow effectively improve the detection precision. |
