Geometry Aware Convolutional Filters for Omnidirectional Images Representation

Pascal Frossard, Renata Khasanova
2019
Article de conférence

Résumé

Due to their wide field of view, omnidirectional cameras are frequently used by autonomous vehicles, drones and robots for navigation and other computer vision tasks. The images captured by such cameras, are often analyzed and classified with techniques designed for planar images that unfortunately fail to properly handle the native geometry of such images and therefore results in suboptimal performance. In this paper we aim at improving popular deep convolutional neural networks so that they can properly take into account the specific properties of omnidirectional data. In particular we propose an algorithm that adapts convolutional layers, which often serve as a core building block of a CNN, to the properties of omnidirectional images. Thus, our filters have a shape and size that adapt to the location on the omnidirectional image. We show that our method is not limited to spherical surfaces and is able to incorporate the knowledge about any kind of projective geometry inside the deep learning network. As depicted by our experiments, our method outperforms the existing deep neural network techniques for omnidirectional image classification and compression tasks.

Official source

https://infoscience.epfl.ch/record/268679?ln=fr

À propos de ce résultat

Cette page est générée automatiquement et peut contenir des informations qui ne sont pas correctes, complètes, à jour ou pertinentes par rapport à votre recherche. Il en va de même pour toutes les autres pages de ce site. Veillez à vérifier les informations auprès des sources officielles de l'EPFL.

Concepts associés

Chargement

Publications associées

Chargement

Though deep learning (DL) algorithms are very powerful for image processing tasks, they generally require a lot of data to reach their full potential. Furthermore, there is no straightforward way to impose various properties, given by the prior knowledge about the target task, on the features extracted by a DL model. Therefore, in this thesis we propose several techniques that rely on the power of graph representations to embed prior knowledge inside the learning process. This allows to reduce the solution space and leads to faster optimization convergence and higher accuracy in the representation learning. In our first work, inspired by the ability of a human to correctly classify rotated, shifted or flipped objects, we propose an algorithm that permits to inherently encode invariance to isometric transformations of objects in an image. Our DL architecture is based on graph representations and consists of three novel layers, which we refer to as graph convolutional, dynamic pooling and statistical layers. Our experiments on the image classification tasks show that our network correctly recognizes isometrically transformed objects even though such types of transformation are not seen by the network at training time. Standard DL techniques are typically not able to succeed in solving such a problem without extensive data augmentation. Then, we propose to exploit the properties of graph-based approaches to efficiently process images with various types of projective geometry. In particular, we are interested in increasingly popular omnidirectional cameras, which have a 360 degree field of view. Despite their effectiveness, such cameras create images with specific geometric properties, which require special techniques for efficient processing. We propose an efficient way of adjusting the weights of the graph edges to adapt the filter responses to the geometric image properties introduced by omnidirectional cameras. Our experiments prove that using the proposed graph with properly adjusted edge weights permits to reach better performance as compared to using regular grid graph with equal weights. Finally, the approach described above relies on the isotropic filters, which work well within our transformation invariant architecture for image classification. However, for other problems (e.g. image compression) or even when used without dynamic pooling and statistical layers that are defined within the proposed architecture, these filters are unable to efficiently encode the information about the object. Thus, we introduce a different technique based on anisotropic filters that adapt their shape and size according to the omnidirectional image geometry. The main advantage of this approach compared to the previous one is the ability to encode the orientation of an image pattern, which is important for various tasks such as image compression. Our experiments show that our approach adapts to different image projective geometries and achieves state-of-the-art performance on image classification and compression tasks. Overall we propose several methods, which combine the power of DL and graph signal processing towards incorporating prior information about the target task inside the optimization procedure. We hope that the research efforts presented in this thesis will help the development of efficient DL algorithms that can use various types of prior knowledge to make them efficient even when the available training data is scarce.

EPFL2019

Chargement

On unifying sparsity and geometry for image-based 3D scene representation

Ivana Tosic

Demand has emerged for next generation visual technologies that go beyond conventional 2D imaging. Such technologies should capture and communicate all perceptually relevant three-dimensional information about an environment to a distant observer, providing a satisfying, immersive experience. Camera networks offer a low cost solution to the acquisition of 3D visual information, by capturing multi-view images from different viewpoints. However, the camera's representation of the data is not ideal for common tasks such as data compression or 3D scene analysis, as it does not make the 3D scene geometry explicit. Image-based scene representations fundamentally require a multi-view image model that facilitates extraction of underlying geometrical relationships between the cameras and scene components. Developing new, efficient multi-view image models is thus one of the major challenges in image-based 3D scene representation methods. This dissertation focuses on defining and exploiting a new method for multi-view image representation, from which the 3D geometry information is easily extractable, and which is additionally highly compressible. The method is based on sparse image representation using an overcomplete dictionary of geometric features, where a single image is represented as a linear combination of few fundamental image structure features (edges for example). We construct the dictionary by applying a unitary operator to an analytic function, which introduces a composition of geometric transforms (translations, rotation and anisotropic scaling) to that function. The advantage of this approach is that the features across multiple views can be related with a single composition of transforms. We then establish a connection between image components and scene geometry by defining the transforms that satisfy the multi-view geometry constraint, and obtain a new geometric multi-view correlation model. We first address the construction of dictionaries for images acquired by omnidirectional cameras, which are particularly convenient for scene representation due to their wide field of view. Since most omnidirectional images can be uniquely mapped to spherical images, we form a dictionary by applying motions on the sphere, rotations, and anisotropic scaling to a function that lives on the sphere. We have used this dictionary and a sparse approximation algorithm, Matching Pursuit, for compression of omnidirectional images, and additionally for coding 3D objects represented as spherical signals. Both methods offer better rate-distortion performance than state of the art schemes at low bit rates. The novel multi-view representation method and the dictionary on the sphere are then exploited for the design of a distributed coding method for multi-view omnidirectional images. In a distributed scenario, cameras compress acquired images without communicating with each other. Using a reliable model of correlation between views, distributed coding can achieve higher compression ratios than independent compression of each image. However, the lack of a proper model has been an obstacle for distributed coding in camera networks for many years. We propose to use our geometric correlation model for distributed multi-view image coding with side information. The encoder employs a coset coding strategy, developed by dictionary partitioning based on atom shape similarity and multi-view geometry constraints. Our method results in significant rate savings compared to independent coding. An additional contribution of the proposed correlation model is that it gives information about the scene geometry, leading to a new camera pose estimation method using an extremely small amount of data from each camera. Finally, we develop a method for learning stereo visual dictionaries based on the new multi-view image model. Although dictionary learning for still images has received a lot of attention recently, dictionary learning for stereo images has been investigated only sparingly. Our method maximizes the likelihood that a set of natural stereo images is efficiently represented with selected stereo dictionaries, where the multi-view geometry constraint is included in the probabilistic modeling. Experimental results demonstrate that including the geometric constraints in learning leads to stereo dictionaries that give both better distributed stereo matching and approximation properties than randomly selected dictionaries. We show that learning dictionaries for optimal scene representation based on the novel correlation model improves the camera pose estimation and that it can be beneficial for distributed coding.

EPFL2009

Chargement

Graph-Based Classification of Omnidirectional Images

Pascal Frossard, Renata Khasanova

Omnidirectional cameras are widely used in such areas as robotics and virtual reality as they provide a wide field of view. Their images are often processed with classical methods, which might unfortunately lead to non-optimal solutions as these methods are designed for planar images that have different geometrical properties than omnidirectional ones. In this paper we study image classification task by taking into account the specific geometry of omnidirectional cameras with graph-based representations. In particular, we extend deep learning architectures to data on graphs; we propose a principled way of graph construction such that convolutional filters respond similarly for the same pattern on different positions of the image regardless of lens distortions. Our experiments show that the proposed method outperforms current techniques for the omnidirectional image classification problem.

2017

Chargement

EPFL2019

On unifying sparsity and geometry for image-based 3D scene representation

Ivana Tosic

EPFL2009