Research On Weakly Supervised Object Detection And Classification Via Multiple Instance Learning Methods

Object detection and classification are two hot research topics in computer vision. Among all the existing research achievements, well labeled images are important preconditions in achieving good performance. However, human annotation is very time-consuming and needs great effort. Whether we can construct a object classification model from the large amount of online images is a very difficult research topic. In this paper, we address the problem of learning object class models from weakly labeled training images, where labels of object classes are only provided at the image level. Weakly supervised object learning is naturally cast as a Multiple Instance Learning(MIL) problem. How to discovery pure positive instances from corrupted data and apply these instances to train a discriminative classifier are challenging problems. We try to address these problems. Specifically, the methods and achievements of the dissertation are as follows.First, we proposed a novel low-rank and sparse constraint based subspace model. Since object instances of a common category are visually similar and when characterized as high-dimensional feature representations, they approximately lie in a low-dimensional subspace. Therefore, we preferred to learn a subspace based generative model for the task of weakly supervised object class learning. Our proposed formulation optimizes a labeling variable for each positive image and learns the subspace model by minimizing rank(via convex surrogate function) of the coefficient matrix associated with the subspace model. To improve discriminative power, our formulation also promotes incoherence betwee n the subspace model and some hard negative instances, which is realized by an ?-insensitive loss. We resort to block coordinate descent and Alternating Direction Method of Multipliers to get locally optimal solutions.Second, we proposed a subspace model based discriminative instances selection(SMDIS) approach. The SMDIS first selects positive instances and then combines with the state-of-the-art Smooth Latent Support Latent Vector Machine(SLSVM) to train a classifier. The SLSVM can perform very well on object detection and classification. With the representative instances selected by the subspace model combined with a powerful SLSVM, we can further improve discriminative ability of the new model. Comprehensive experiments have been performed on several face data sets, i.e., AR, UMIST, ORL data sets and PASCAL 2007 data sets, which demonstrate that our method achieves comparably better results compared with other subspace learning approaches.In most cases of the real world, more than one instance is included in one image. In fact, we can build a strong model by ensemble multiple subspace models with Ada Boost strategy. The ensemble model learns a set of weights assigned to each classifier and considers advantages of each single model as well as focus on the difficult classified samples. By the proposed ensemble model, the detection results on Yale B and AR face detection data sets are quite promising which outperformed when used only one single subspace model.