Surface representations and automatic feature learning for 3D object recognition

Vision-based object recognition is a popular area that has gained a significant popularity in the last decade because of its many applications including robotics, medical, manufacturing and video surveillance. The aim of object recognition is to identify object sin a scene and estimate their pose. An important paradigm of object recognition is to first define suitable surface representations, offline, for 3D objects and save those representations in a database. During online recognition, a similar representation of a scene is matched with the representations stored in the database to recognize objects which are present in the scene. The main challenges associated with surface representation and 3D object recognition are occlusions caused by the presence of multiple objects in the scene, clutter due to unwanted objects, and robustness to noise and resolution.

This dissertation addresses the aforementioned challenges and investigates novel surface representations, matching and automatic feature learning techniques for object recognition. In this thesis, three different surface representations, a keypoint detector, automatic feature matching and recognition algorithms as well as two different automatic feature learning techniques are presented.

The first part of the thesis presents two different local surface representations based on the analysis of the 3D vector field. The divergence and vorticity of the vector field have been exploited to construct two different features named 3D-Div and 3D-Vor. To achieve invariance to rigid transformations, each representation is defined on an object centered local reference frame. In addition to the aforementioned surface representations, this dissertation also presents a keypoint-based feature-free representation for 3D modeling and object recognition. Instead of computing the local feature around a given 3D keypoint, the proposed technique measures the geometrical relationships between the de-tected keypoints for surface representation. The proposed representation is found to be computationally efficient compared to existing feature based methods. This thesis also presents novel algorithms for range image registration, 3D keypoint detection and 3D object recognition in complex scenes in the presence of occlusions and clutter. The presented methods achieve superior performance compared to existing techniques when tested on publicly available datasets including the low resolution Washington RGB-D object, UWA,Bologna, Stanford 3D model and Ca’Foscari datasets.

The last part of this dissertation investigates two novel automatic feature learning techniques through extensive experiments using publicly available object/face datasets. In the first technique, an Iterative Deep Learning Model (IDLM) is presented. IDLM consists of pooled convolutional layer followed by artificial neural networks applied in a hierarchical fashion to automatically learn discriminative representations from raw face and object images. The proposed deep learning framework is extensively tested on four publicly available object and face datasets and achieves superior performance compared to existing methods. In the second proposed technique, dubbed Evolutionary Feature Learning (EFL), evolutionary algorithms are exploited for 3D object recognition. In the EFL process, irrelevant and redundant features are omitted: only the features that describe theobject in the most discriminative manner are selected during the evolutionary process.The proposed technique automatically optimizes the candidate solution based on the fitnessfunction and selects the best feature for superior object recognition. The proposedapproach was extensively evaluated on four challenging object datasets achieving superiorperformance compared to existing techniques including deep learning and dictionarylearning based methods.