Karla Brkić

A review of feature selection methods with applications

Feature selection (FS) methods can be used in data pre-processing to achieve efficient data reduction. This is useful for finding accurate data models. Since exhaustive search for optimal feature subset is infeasible in most cases, many search strategies have been proposed in literature. The usual applications of FS are in classification, clustering, and regression tasks. This review considers most of the commonly used FS techniques. Particular emphasis is on the application aspects. In addition to standard filter, wrapper, and embedded methods, we also provide insight into FS for recent hybrid approaches and other advanced topics.

Exploiting temporal and spatial constraints in traffic sign detection from a moving vehicle

This paper addresses detection, tracking and recognition of traffic signs in video. Previous research has shown that very good detection recalls can be obtained by state-of-the-art detection algorithms. Unfortunately, satisfactory precision and localization accuracy are more difficultly achieved. We follow the intuitive notion that it should be easier to accurately detect an object from an image sequence than from a single image. We propose a novel two-stage technique which achieves improved detection results by applying temporal and spatial constraints to the occurrences of traffic signs in video. The first stage produces well-aligned temporally consistent detection tracks by managing many competing track hypotheses at once. The second stage improves the precision by filtering the detection tracks by a learned discriminative model. The two stages have been evaluated in extensive experiments performed on videos acquired from a moving vehicle. The obtained experimental results clearly confirm the advantages of the proposed technique.

Generative modeling of spatio-temporal traffic sign trajectories

We consider the task of automatic detection and recognition of traffic signs in video. We show that successful off-the-shelf detection (Viola-Jones) and classification (SVM) systems yield unsatisfactory results. Our main concern are high false positive detection rates which occur due to sparseness of the traffic signs in videos. We address the problem by enforcing spatio-temporal consistency of the detections corresponding to a distinct sign in video. We also propose a generative model of the traffic sign motion in the image plane, which is obtained by clustering the trajectories filtered by an appropriate procedure. The contextual information recovered by the proposed model will be employed in our future research on recognizing traffic signs in video.

A computer vision assisted geoinformation inventory for traffic infrastructure

Geoinformation inventories are often employed as a tool for providing a comprehensive view onto the required state of traffic control infrastructure. They are especially important in road safety inspection where, in combination with georeferenced video, they enable repeatable off-line and off-site assessments as an attractive aternative to classic onsite inspection. Nevertheless, manual assessments are tedious and time-consuming even when performed off-line, and this seriously impairs the potential of the geoinformation inventory concept. This paper therefore researches a hypothesis that suitable georeferenced video processing techniques would allow reliable automation of the following operations: i) creation of the traffic inventory from the given video, and ii) assessing the video against the state in the inventory. Prominent computer vision approaches have been rigorously and systematically evaluated and the obtained results are presented. The results seem to support the hypothesis, although further work is required for a more definite answer.

Decision Tree Ensembles in Biomedical Time-Series Classification

There are numerous classification methods developed in the field of machine learning. Some of these methods, such as artificial neural networks and support vector machines, are used extensively in biomedical time-series classification. Other methods have been used less often for no apparent reason. The aim of this work is to examine the applicability of decision tree ensembles as strong and practical classification algorithms in biomedical domain. We consider four common decision tree ensembles: AdaBoost.M1+C4.5, Multi- Boost+C4.5, random forest, and rotation forest. The decision tree ensembles are compared with SMO-based support vector machines classifiers (linear, squared polynomial, and radial kernel) on three distinct biomedical time-series datasets. For evaluation purposes, 10x10-fold cross-validation is used and the classifiers are measured in terms of sensitivity, specificity, and speed of model construction. The classifiers are compared in terms of statistically significant winslosses-ties on the three datasets. We show that the overall results favor decision tree ensembles over SMO-based support vector machines. Preliminary results suggest that AdaBoost.M1 and MultiBoost are the best of the examined classifiers, with no statistically significant difference between them. These results should encourage the use of decision tree ensembles in biomedical time-series datasets where optimal model accuracy is sought.

Image representations on a budget: Traffic scene classification in a restricted bandwidth scenario

Modern fleet management systems typically monitor the status of hundreds of vehicles by relying on GPS and other simple sensors. Such systems experience significant problems in cases of GPS glitches as well as in areas without GPS coverage. Additionally, when the tracked vehicle is stationary, they cannot discriminate between traffic jams, service stations, parking lots, serious accidents and other interesting scenarios. We propose to alleviate these problems by augmenting the GPS information with a short descriptor of an image captured by an on-board camera. The descriptor allows the server to recognize various scene types by image classification and to subsequently implement suitable business policies. Due to restricted bandwidth we focus on finding a compact image representation that would still allow reliable classification. We therefore consider several state-of-the-art descriptors under tight representation budgets of 512, 256, 128 and 64 components, and evaluate classification performance on a novel image dataset specifically crafted for fleet management applications. Experimental results indicate fair performance even with very short descriptor sizes and encourage further research in the field.

I Know That Person: Generative Full Body and Face De-identification of People in Images

We propose a model for full body and face deidentification of humans in images. Given a segmentation of the human figure, our model generates a synthetic human image with an alternative appearance that looks natural and fits the segmentation outline. The model is usable with various levels of segmentation, from simple human figure blobs to complex garment-level segmentations. The level of detail in the de-identified output depends on the level of detail in the input segmentation. The model de-identifies not only primary biometric identifiers (e.g. the face), but also soft and non-biometric identifiers including clothing, hairstyle, etc. Quantitative and perceptual experiments indicate that our model produces de-identified outputs that thwart human and machine recognition, while preserving data utility and naturalness.

Towards neural art-based face de-identification in video data

We propose a computer vision-based pipeline that enables altering the appearance of faces in videos. Assuming a surveillance scenario, we combine GMM-based background subtraction with an improved version of the GrabCut algorithm to find and segment pedestrians. Independently, we detect faces using a standard face detector. We apply the neural art algorithm, utilizing the responses of a deep neural network to obfuscate the detected faces through style mixing with reference images. The altered faces are combined with the original frames using the extracted pedestrian silhouettes as a guideline. Experimental evaluation indicates that our method has potential in producing de-identified versions of the input frames while preserving the utility of the de-identified data.

Iterative Automated Foreground Segmentation in Video Sequences Using Graph Cuts

In this paper we propose a method for foreground object segmentation in videos using an improved version of the GrabCut algorithm. Motivated by applications in de-identification, we consider a static camera scenario and take into account common problems with the original algorithm that can result in poor segmentation. Our improvements are as follows: (i) using background subtraction, we build GMM-based segmentation priors; (ii) in building foreground and background GMMs, the contributions of pixels are weighted depending on their distance from the boundary of the object prior; (iii) probabilities of pixels belonging to foreground or background are modified by taking into account the prior pixel classification as well as its estimated confidence; and (iv) the smoothness term of GrabCut is modified by discouraging boundaries further away from the object prior. We perform experiments on CDnet 2014 Pedestrian Dataset and show considerable improvements over a reference implementation of GrabCut.

Automatic recognition of handwritten corrections for multiple-choice exam answer sheets

Automated grading of multiple-choice exams is of great interest in university courses with a large number of students. We consider an existing system in which exams are automatically graded using simple answer sheets that are annotated by the student. A sheet consists of a series of circles representing possible answers. As annotation errors are possible, a student is permitted to alter the annotated answer by annotating the“error” circle and handwriting the letter of the correct answer next to the appropriate row. During the scanning process, if an annotated“error” circle is detected, the system raises an alarm and requires intervention from a human operator to determine which answer to consider valid. We propose rather simple and effecive computer vision algorithm which enables automated reading of a limited set of handwritten answers and minimizes the need for a human intervention in the scanning process. We test our algorithm on a large dataset of real scanned answer sheets, and report encouraging performance rates.