Pekka Astola

Sparse modelling and predictive coding of subaperture images for lossless plenoptic image compression

This paper studies the lossless compression of rectified light-field images captured by plenoptic cameras, exploiting the high similarity existing between the subaperture images, or views, composing the light-field image. The encoding is predictive, where one sparse predictor is designed for every region of a view, using as regressors the pixels from the already transmitted views. As a first step, consistent segmentations for all subaperture images are constructed, defining the regions as connected components in the quantized depth map of the central view, and then propagating them to all side views. The sparse predictors are able to take into account the small horizontal and vertical disparities between regions in corresponding close-by views and perform optimal least squares interpolation accounting implicitly for fractional disparities. The optimal structure of the sparse predictor is selected for each region based on an implementable description length. The encoding of the views is done sequentially starting from the central view and the scheme produces results better than standard lossless compression methods utilized directly on the full lightfield image or applied to the views in a similar sequential order as our method.

Sparse prediction for compression of stereo color images conditional on constant disparity patches

This paper introduces an algorithm for lossless encoding of color stereo images using sparse prediction and context coding. For encoding the left color image, an extension of the method uses additionally conditioning on the warped image, obtained by warping the right color image using the information in the disparity image. Different sparse predictors are designed and used at the locations of large constant patches of the disparity image. The new method of stereo color image compression is shown to perform better than several publicly available lossless image compressors over the Middlebury dataset. Besides compression applications, the method can be used for computing an implementable minimum description length cost for ranking candidate disparity map images, according to their performance in the warping process.

Minimum Description Length Sparse Modeling and Region Merging for Lossless Plenoptic Image Compression

This paper proposes a complete lossless compression method for exploiting the redundancy of rectified light-field data. The light-field data consist of an array of rectified subaperture images, called for short views, which are segmented into regions according to an optimized partition of the central view. Each region of a view is predictively encoded using a specifically designed sparse predictor, exploiting the smoothness of each color component in the current view, and the cross similarities with the other color components and already encoded neighbor views. The views are encoded sequentially, using a spiral scanning order, each view being predicted based on several similar neighbor views. The essential challenge for each predictor becomes choosing the most relevant regressors from a large number of possible regressors belonging to the neighbor views. The proposed solution here is to couple sparse predictor design and minimum description length (MDL) principle, where the data description length is measured by an implementable code length, optimized for a class of probability models. This paper introduces a region merging segmentation under the MDL criterion for partitioning the views into regions having their own specific sparse predictors. In experiments, several fast sparse design methods are considered. The proposed scheme is evaluated over a database of plenoptic images, achieving better lossless compression ratios than straightforward usage of standard image and video compression methods for the spiral sequence of views.

Precise outline matching criteria for target pose estimation and odometry from stereo video

We introduce and investigate several criteria for matching the outer contour generated by a model of the manipulation target against the image data captured with a stereo camera rig. The more complex criteria are intended for the initial stage of scene analysis, to provide pose and location estimation for path planning, and the faster methods can be used in real-time, e.g. for providing feedback during visual servoing. We exemplify the methodology for the case of cylinder pose estimation, which is an often analyzed case in the literature, and we first investigate the pose estimation performance using simulated images, where the ground truth is perfectly known. Experimental results using real video sequences captured from the mobile manipulator iMoro are showing very good matching for target pose estimation, in agreement with results on simulated data. The visual odometry results using the pose estimation over successive frames show a smooth path estimation and a simulation study infers small localization errors along the path.

Lossless compression of high resolution disparity map images

High resolution disparity images are stored in floating point raw files, where the number of bits per pixel is typically 32, although the number of used bits when converted to a fixed point representation is lower, e.g., between 24 and 26 in the dataset used in our experiments. In order to compress images with such high dynamic range, the bitplanes of the original image are combined into integer images with at most 16 bits, for which readily existing compressors are available. We introduce first a context predictive compressor (CPC) which can operate on integer images having more than 16 bits. The proposed overall compression scheme uses a first revertible linear transformation of the image as a first decorrelation process, and then splits the transformed image into integer images with smaller dynamic range, which are finally encoded. We experiment with schemes of split-into-2 and split-into-3, with combinations of several existing compressors for the integer image components and show that the newly introduced CPC operating over the least significant bitplanes combined with CERV operating over the most significant bitplanes achieves always the best compression, with final lossless compressed results of between 8 and 12 bits per pixel.

Algebraic and Combinatorial Methods for Reducing the Number of Variables of Partially Defined Discrete Functions

Applications of pattern recognition, design of faulttolerant systems and communications have key problems that arenaturally described by partially defined (incompletely defined)discrete functions. Such partially defined functions arising frompractical demands usually have a large number of variables andso their direct implementations require complex systems. Thusit is important to have at hand an efficient method to reducethe number of their variables. Here we review recent results tolinearly decompose a discrete function using a transform thatcan be efficiently implemented as a Galois field deconvolution. We also study the question: What are the general bounds for thedimension of the range space for an arbitrary linear transformto reduce a partially defined discrete function? We derive abound for the dimension of the range for arbitrary lineartransformation. We also estimate how good linear decompositioncan be obtained by the use of random transformations and showthat with a randomly generated transform we can reach theabove discussed bound.

Object detection in robotic applications for real-time localization using semi-unknown objects

We investigate the pose estimation of a semi-unknown object for stereo-vision-based navigation of a mobile manipulator. A new computationally fast vision algorithm is developed to extract the objects pose at a high rate from the captured scenes. Moreover, we present a method to deal with range dependent noise characteristics of the stereo vision to fulfill requirements for mobile manipulation tasks. As shown, a smoothed, high-bandwidth feedback is obtained by using robust real-time estimation, where special care is taken to accommodate the aforementioned nonlinearities of the stereo vision. This way, the manipulator is capable of positioning itself in the close vicinity of an object by navigation of its nonholonomic mobile base. Importantly, we achieve nearly the same accuracy in mobile robot positioning compared to standard marker-based techniques at distances greater than those typically considered suitable for position-based, high-bandwidth motion control within the robotics community.

Immersion depth estimation using spectrograms displaying Lloyd's mirror patterns

Lloyds mirror effect has been used in the past for estimating the location parameters of an underwater target, under the hypothesis of constant speed of sound, in which the sound propagates along straight lines. However, as the sound speed depends on the water temperature, which typically varies with depth, the sound propagation paths become curved. In this paper, we use a ray tracing method for computing the time delays along the direct and the reflected signals under different sound speed profiles in a shallow water environment. The time delays are used to synthesize a sensor signal containing multipath interference. The location parameter of the target can be estimated by matching a measured spectrogram against libraries of simulated spectrograms. In the experiments we illustrate the estimation process and the robustness against uncertainty in the sound speed profiles among the libraries.

Lossless compression of regions-of-interest from retinal images

This paper presents a lossless compression method performing separately the compression of the vessels and of the remaining part of eye fundus in retinal images. Retinal images contain valuable information sources for several distinct medical diagnosis tasks, where the features of interest can be e.g. the cotton wool spots in the eye fundus, or the volume of the vessels over concentric circular regions. It is assumed that one of the existent segmentation methods provided the segmentation of the vessels. The proposed compression method transmits losslessly the segmentation image, and then transmits the eye fundus part, or the vessels image, or both, conditional on the vessels segmentation. The independent compression of the two color image segments is performed using a sparse predictive method. Experiments are provided over a database of retinal images containing manual and estimated segmentations. The codelength of encoding the overall image, including the segmentation and the image segments, proves to be better than the codelength for the entire image obtained by JPEG2000 and other publicly available compressors.