Olga Barinova

On Detection of Multiple Object Instances Using Hough Transforms

Hough transform-based methods for detecting multiple objects use nonmaxima suppression or mode seeking to locate and distinguish peaks in Hough images. Such postprocessing requires the tuning of many parameters and is often fragile, especially when objects are located spatially close to each other. In this paper, we develop a new probabilistic framework for object detection which is related to the Hough transform. It shares the simplicity and wide applicability of the Hough transform but, at the same time, bypasses the problem of multiple peak identification in Hough images and permits detection of multiple objects without invoking nonmaximum suppression heuristics. Our experiments demonstrate that this method results in a significant improvement in detection accuracy both for the classical task of straight line detection and for a more modern category-level (pedestrian) detection problem.

On detection of multiple object instances using hough transforms

Hough transform-based methods for detecting multiple objects use nonmaxima suppression or mode seeking to locate and distinguish peaks in Hough images. Such postprocessing requires the tuning of many parameters and is often fragile, especially when objects are located spatially close to each other. In this paper, we develop a new probabilistic framework for object detection which is related to the Hough transform. It shares the simplicity and wide applicability of the Hough transform but, at the same time, bypasses the problem of multiple peak identification in Hough images and permits detection of multiple objects without invoking nonmaximum suppression heuristics. Our experiments demonstrate that this method results in a significant improvement in detection accuracy both for the classical task of straight line detection and for a more modern category-level (pedestrian) detection problem.

Fast Automatic Single-View 3-d Reconstruction of Urban Scenes

We consider the problem of estimating 3-d structure from a single still image of an outdoor urban scene. Our goal is to efficiently create 3-d models which are visually pleasant. We chose an appropriate 3-d model structure and formulate the task of 3-d reconstruction as model fitting problem. Our 3-d models are composed of a number of vertical walls and a ground plane, where ground-vertical boundary is a continuous polyline. We achieve computational efficiency by special preprocessing together with stepwise search of 3-d model parameters dividing the problem into two smaller sub-problems on chain graphs. The use of Conditional Random Field models for both problems allows to various cues. We infer orientation of vertical walls of 3-d model vanishing points.

Image Binarization for End-to-End Text Understanding in Natural Images

While modern off-the-shelf OCR engines show particularly high accuracy on scanned text, text detection and recognition in natural images still remains a challenging problem. Here, we demonstrate that OCR engines can still perform well on this harder task as long as appropriate image binarization is applied to input photographs. For such binarization, we systematically evaluate the performance of 12 binarization methods as well as of a new binarization algorithm that we propose here. Our evaluation includes different metrics and uses established natural image text recognition benchmarks (ICDAR 2003 and ICDAR 2011). Our main finding is thus the fact that image binarization methods combined with additional filtering of generated connected components and off-the-shelf OCR engines can achieve state-of-the-art performance for end-to-end text understanding in natural images.

Fast and accurate scene text understanding with image binarization and off-the-shelf OCR

While modern off-the-shelf OCR engines show particularly high accuracy on scanned text, text detection and recognition in natural images still remain a challenging problem. Here, we demonstrate that OCR engines can still perform well on this harder task as long as an appropriate image binarization is applied to input photographs. We propose a new binarization algorithm that is particularly suitable for scene text and systematically evaluate its performance along with 12 existing binarization methods. While most existing binarization techniques are designed specifically either for text detection or for recognition of localized text, our method shows very similar results for both large images and localized text regions. Therefore, it can be applied to large images directly with no need for re-binarization of localized text regions. We also propose the real-time variant of this method based on linear-time bilateral filtering. Evaluation across different metrics on established natural image text recognition benchmarks (ICDAR 2003 and ICDAR 2011) shows that our simple and fast image binarization method combined with off-the-shelf OCR engine achieves state-of-the-art performance for end-to-end text understanding in natural images and outperforms recent fancy methods.

Learning graph laplacian for image segmentation

In this paper we formulate the task of semantic image segmentation as a manifold embedding problem and solve it using graph Laplacian approximation. This allows for unsupervised learning of graph Laplacian parameters individually for each image without using any prior information. We perform experiments on GrabCut, Graz and Pascal datasets. At a low computational cost proposed learning method shows comparable performance to choosing the parameters on the test set. Our framework for semantic image segmentation shows better performance than the standard discrete CRF with graph-cut inference.

ODDboost: Incorporating Posterior Estimates into AdaBoost

Boosting methods while being among the best classification methods developed so far, are known to degrade performance in case of noisy data and overlapping classes. In this paper we propose a new upper generalization bound for weighted averages of hypotheses, which uses posterior estimates for training objects and is based on reduction of binary classification problem with overlapping classes to a deterministic problem. If we are given accurate posterior estimates, proposed bound is lower than existing bound by Schapire et al [25]. We design an AdaBoost-like algorithm which optimizes proposed generalization bound and show that incorporated with good posterior estimates it performs better than the standard AdaBoost on real-world data sets.

Generic Regularization of Boosting-Based Algorithms for the Discovery of Regime-Independent Portfolio Strategies from High-Noise Time Series

Recently proposed boosting-based optimization offers a generic framework for the discovery of portfolios of complementary base trading strategies with stable combined performance over wide range of market regimes and robust generalization abilities. However, wide variety of market regimes and existence of hard-to-model periods reduces universe of financial instruments and achievable performance ranges for which such portfolio strategies can be found. Recently introduced generic regularization approach based on confusing (noisy) sample removal was shown to be effective for diversification of portfolio strategies discovered by boosting-based optimization. Here we argue and demonstrate that this regularization technique could be also effective in dealing with large periods of excessive volatility and significantly reduced determinism in training data. In the most recent history such situation occurred during current financial crisis. The algorithm for confusing sample removal is outlined and applied to the recent market data in the context of mid-frequency intraday trading.