The Research On Part-based Methods For Non-rigid Object Detection

Object detection is a classic problem in object recognition and is viewed as the basis of several high-level tasks.The research on feature description and detection methods embeds the overall developments in recognition technology.Among the object categories,the detection of non-rigid objects is more complex comparing other objects.The main challenges from non-rigid objects include a huge variation of intra-class appearance,severe noise affecting feature learning,and local deformation.Part-based models provide a powerful description method for object categories,especially non-rigid objects.In these models an object is represented by a collection of parts arranged in a deformable configuration.In this configuration pairs of parts are modeled as spring-like spatial connections,and the appearance of each part is described separately.Part models realize the divide-and-conquer methods for detecting a whole object through finding its parts.This paper gives an overview of current part-based methods and models,and analyzes the fundamentals and key techniques of them.And then we conclude a clear definition of parts and part-based models under the background of the detection task.Forcusing on the non-rigid objects,this paper proposes several new methods and models in order to improve the detection performance.The research route is:firstly,handling the drawbacks of classic detection algorithms by applying part-based methods;secondly,improving current part models from multiple aspects.All experiments in this paper are based on standard and published datasets.We compare the proposed algorithms and models with current state-of-the-art methods.The experimental results illustrate the effectiveness of our methods and give some valuable conclusions.The main works and innovations are given as follows.1.A part-based voting method is propsed to address the high rates of false positives in probabilistic voting methods.Probabilistic voting models are important methods for detecting non-rigid objects as well as recognizing their contours.There are two main factors that affect the detection performance of voting methods.One factor is that the voting elements are collected and are trained based on naive methods that are sensitive to noise.The other factor is that image features vote for an object position independently,resulting in votes for object positions in low reliability.We propose a discriminative voting method based on parts.The proposed method measures a voting score of an object position in a discriminative way rather than a probabilistic way,which supports a margin-based optimization for the learning of element weights and improves the learning performance.On the other hand,the method groups voting elements into parts and casts consistent votes for one object position based on the dependences of local features that belong to the same part.The consistent votes have more reliable than independent votes and they can reduce error votes from noise.2.An adaptive part learning method is proposed to address the problem of feature learning in weakly-supervision.Current part models,although providing a feature learning method by automatically localizing discriminative parts with maximum energy,cannot ensure that the localized parts are relevant to the object in a weakly-supervised setting.We consider a fusion of multiple cues to localize both discriminative and semantically meaningful parts in an adaptive way.The key idea is to measure consistency on the distribution on color,texture and edge for every candidate part.The higher consistency a part has on the distribution,the more relevant it is to the object.We propose two part localization schemes:one is to set the adaptive method as the post-processing of previous searching methods;the other is to integrate the search and localization by constructing a shape-based part model.When a training example has been partially occluded,the parts that are selected from the occluded area are actually noise.To reducing the noise,we propose a pruning scheme during the part learning,which evaluate whether a part sample is false positive or not according to its classification score.The proposed scheme can effectively reduce the ratio of noise and enhance the convergence rate.3.For the problem of intra-class variations in articulated objects,a hierarchical model is proposed by jointing object detection and pose estimation.Part models usually classify the various appearance of the object according to aspect-ratio of training examples.This kind of classification cannot reflect the real appearance distribution of non-rigid objects because the changes of the appearance mainly arise from different poses.In the hierarchical model,the pose of the object is recognized and is considered as an evidence for detecting the object.The proposed model has two characteristics.The first one is that it realizes object detection and pose estimation in parallel,addressing the error propagation problem existing in current joint models;the second one is that it balances the performance of detection and estimation.4.Automatic part annotation could provide part-level training examples for the supervised learning of part models.However,the different feature distribution on soure-target datasets is usually faced under the real-world condition,which results in a limitation on learning performance for annotation models when traditional mechine learning methods are used.This paper proposes a domain adaption algorithm coping with the problem and applies it to part annotation task.The proposed algorithm modifies the loss function of a traditional structural SVM by defining a new regularizer based on parameter similarity.The objective is to minimize the optimization loss and to find the target parameters most similar to source ones at the same time.Aiming at improving learning performance,two active learning strategies are integrated into the method in order to choose 'high quality' samples for training.Experiments include two adaptation scenarios.That is,the training samples between source-target datasets have different pose distribution or different appearance distribution.Experimental results show the integrated algorithm outperforms the traditional structural SVM obviously.