Richard Rzeszutek

Semi-automatic 2D to 3D image conversion using scale-space Random Walks and a graph cuts based depth prior

In this paper, we present a semi-automated method for converting conventional 2D images into stereoscopic 3D. User-defined strokes corresponding to a rough estimate of the depth values in the scene are defined for the image of interest. With these, our system determines the depth values for the rest of the image, producing a depth map that can be used to create stereoscopic 3D image pairs. Our work is based on a similar scheme, using the Random Walks segmentation paradigm. However, the related work is quite complex, with many processing steps required to produce the final stereoscopic image pair. Combined with its evident shortcomings, but noting the merits, we propose a system employing Random Walks, while incorporating information from the popular Graph Cuts segmentation paradigm. Thus, a final cohesive depth map is produced, combining the merits of both. The results show that we can produce good quality stereoscopic image pairs, while using a much more simplified method in comparison to the related work.

Semi-automatic synthetic depth map generation for video using random walks

We present a method for easily generating depth maps from monoscopic (i.e. “2D”) video footage in order to convert them into stereoscopic, or “3D”, footage. Our method uses user-defined strokes for a number of keyframes in the original footage and interpolates between the keyframes to provide a sparse labelling for each frame. We then apply the Random Walks algorithm to the footage to provide depth estimates based on the input provided by the user. These depth maps can then be used to generate novel views through depth-based image rendering.

Image segmentation using Scale-Space Random Walks

Many methods for supervised image segmentation exist. One such algorithm, Random Walks, is very fast and accurate when compared to other methods. A drawback to Random Walks is that it has difficulty producing accurate and clean segmentations in the presence of noise. Therefore, we propose an extension to Random Walks that improves its performance without significantly modifying the original algorithm. Our extension, known as “Scale-Space Random Walks”, or SSRW, addresses these problems. The SSRW is able to produce more accurate segmentations in the presence of noise while still retaining all of the properties of the original algorithm.

Semi-automatic 2D to 3D image conversion using a hybrid Random Walks and graph cuts based approach

In this paper, we present a semi-automated method for converting conventional 2D images to stereoscopic 3D. User-defined strokes that correspond to a rough estimate of the depth values in the scene are defined for the image of interest. With these strokes, our system thus determines what the depth values are for the rest of the image, producing a depth map that is ultimately used to create a stereoscopic image pair. Our work is based on a similar scheme which employs Random Walks. However, the related work is quite complex, with many processing steps required to produce the final stereoscopic image pair. Combined with the evident shortcomings of the related work, but noting the merits of Random Walks, we propose a system that is a hybrid between Random Walks, and the popular Graph Cuts segmentation paradigm. Both segmentation algorithms are used to generate a final cohesive depth map, thus combining the merits of both frameworks together. The generated results show that we can produce good quality stereoscopic image pairs, while using a much more simplified method in comparison to the related Random Walks scheme.

Self-Organizing Maps for Topic Trend Discovery

The large volume of data on the Internet makes it extremely difficult to extract high-level information, such as recurring or time-varying trends in document content. Dimensionality reduction techniques can be applied to simplify the analysis process but the amount of data is still quite large. If the analysis is restricted to just text documents then Latent Dirichlet Allocation (LDA) can be used to quantify semantic, or topical, groupings in the data set. This paper proposes a method that combines LDA with the visualization capabilities of Self-Organizing Maps to track topic trends over time. By examining the response of a map over time, it is possible to build a detailed picture of how the contents of a dataset change.

Efficient automatic depth estimation for video

Estimating depth in monoscopic images and videos is a non-trivial problem due to the inherent ambiguity that arises when a 3D scene is projected onto a 2D plane (the image). But because depth estimation is so useful, many different techniques have been developed to solve this problem. Unfortunately these methods tend to be computationally intensive or require precise knowledge about the camera that captured the scene. We present a simple and straightforward technique that can estimate relative depth in video sequences using well-established computer vision principles. We also utilize recent advancements in non-linear filtering to make the estimation process computationally efficient. The result produces depth maps comparable to ground truth depths extracted by state-of-the-art estimation methods.

Interactive rotoscoping through scale-space random walks

We present a novel rotoscoping method that provides accurate object boundaries while still remaining simple to use. We have designed this method so that a user familiar with existing, spline-based rotoscoping techniques will not need to significantly modify their workflow. Our method utilizes a modified version of the Random Walks segmentation algorithm, known as Scale-Space Random Walks (SSRW), to provide a level of accuracy simply not possible using splines alone. Furthermore, due to the nature of the SSRW algorithm, we are able to provide a realtime, or near realtime, level of interaction with the user.

Propagating sparse labels through edge-aware filters

In this paper, we show how linear, but not necessarily shift-invariant, filters can be used to propagate sparse labels throughout an image. We propose a new propagation method based on the domain transform filter, a linear, shift-varying filter whose kernel width varies based on local edge information. We contrast this against the more well-known energy minimization approach and show that it can produce significantly better results when the input labels contain errors. Finally, we show how minimization-based methods are better suited for purely user-guided applications.

A framework for estimating relative depth in video

Abstract We present a method for efficiently generating dense, relative depth estimates from video without requiring any knowledge of the imaging system, either a priori or by estimating it during processing. Instead we only require that the epipolar constraint between any two frames is satisfied and that the fundamental matrix can be estimated. By tracking sparse features across many frames and aggregating the multiple depth estimates together, we are able to improve the overall estimate for any given frame. Once the depth estimates are available, we treat the generation of the depth maps as a label propagation problem. This allows us to combine the automatically generated depth maps with any user corrections and modifications (if so desired).

Label propagation through edge-preserving filters

In this paper we investigate methods for propagating automatically generated or user-defined labels through an image using edge-preserving filters. We focus on the domain transform filter as it has been used for propagation purposes in the past. The method we present addresses some of the numerical issues that arise with using the filter directly and also improve on the results by better respecting the underlying image structure during the label propagation. Finally we also demonstrate how a filter-based approach is preferable to using global optimization for interpolating automatically generated sparse features.