Domonkos István Varga

Robust real-time pedestrian detection in surveillance videos

Detecting different categories of objects in an image and video content is one of the fundamental tasks in computer vision research. Pedestrian detection is a hot research topic, with several applications including robotics, surveillance and automotive safety. We address the problem of detecting pedestrians in surveillance videos. In this paper, we present a new feature extraction method based on Multi-scale Center-symmetric Local Binary Pattern operator. All the modules (foreground segmentation, feature pyramid, training, occlusion handling) of our proposed method are introduced with its details about design and implementation. Experiments on CAVIAR and other sequences show that the presented system can detect pedestrians in real-time effectively and accurately in surveillance videos.

Fast content-based image retrieval using convolutional neural network and hash function

Due to the explosive increase of online images, content-based image retrieval has gained a lot of attention. The success of deep learning techniques such as convolutional neural networks have motivated us to explore its applications in our context. The main contribution of our work is a novel end-to-end supervised learning framework that learns probability-based semantic-level similarity and feature-level similarity simultaneously. The main advantage of our novel hashing scheme that it is able to reduce the computational cost of retrieval significantly at the state-of-the-art efficiency level. We report on comprehensive experiments using public available datasets such as Oxford, Holidays and ImageNet 2012 retrieval datasets.

Fully automatic image colorization based on Convolutional Neural Network

This paper deals with automatic image colorization. This is a very difficult task, since it is an ill-posed problem that usually requires user intervention to achieve high quality. A fully automatic approach is proposed that is able to produce realistic colorization of an input grayscale image. Motivated by the recent success of deep learning techniques in image processing, we propose a feed-forward, two-stage architecture based on Convolutional Neural Network that predicts the U and V color channels. Unlike most of the previous works, this paper presents a fully automatic colorization which is able to produce high-quality and realistic colorization even of complex scenes. Comprehensive experiments and qualitative and quantitative evaluations were conducted on the images of SUN database and on other images. We have found that Quaternion Structural Similarity (QSSIM) gives in some degree a good base for quantitative evaluation, that is why we chose QSSIM as an index-number for the quality of colorization.

Pedestrian detection in surveillance videos based on CS-LBP feature

Detecting different categories of objects in an image and video content is one of the fundamental tasks in computer vision research. Pedestrian detection is a hot research topic, with several applications including robotics, surveillance and automotive safety. Pedestrians are key participants in transportation systems, so pedestrian detection in video surveillance systems is of great significance to the research and application of Intelligent Transportation Systems (ITS). Pedestrian detection is a challenging problem due to the variance of illumination, color, scale, pose, and so forth. Extraction of effictive features is a key to this task. In this work, we present the multi-scale Center-symmetric Local Binary Pattern feature for pedestrian detection. The proposed feature captures gradient information and some texture and scale information. We completed the detection task with a foreground segmentation method. Experiments on CAVIAR sequences show that the proposed feature with support vector machines can detect pedestrians in real-time effectively in surveillance videos.

LHI-tree: An efficient disk-based image search application

Querying of nearest neighbour (NN) elements on large data collections is an important task for several information or content retrieval tasks. In the paper Local Hash-indexing tree (LHI-tree) is introduced, which is a disk-based index scheme that uses RAM for quick space partition localization and hard disks for the hash indexing. When large collections are considered, such hybrid data structure should be used to implement an effective indexing service. The proposed structure can produce list of approximate nearest neighbours. We compare LHI-tree to FLANN (Fast Library for Approximate Nearest Neighbors), an effective and frequently instanced implementation of ANN search. We show that they produce similar lists of retrieved images (although FLANN works only on RAM). In case of huge multimedia database a disk based indexing and retrieval method has a significant advantage against a vector based system running in RAM data. For the visual content indexing we built an image descriptor composed of four different information representations: edge histogram, entropy histogram, pattern histogram and dominant component colour characteristics. The paper will mention the content based retrieval of Hungarian Wikipedia images.

Search in WikiImages using mobile phone

Demonstration will focus on the content based retrieval of Wikipedia images (Hungarian version). A mobile application for iOS will be used to gather images and send directly to the crossmodal processing framework. Searching is implemented in a high performance hybrid index tree with total ~500k entries. The hit list is converted to wikipages and ordered by the content based score.

Person re-identification based on deep multi-instance learning

Person re-identification is one of the widely studied research topic in the fields of computer vision and pattern recognition. In this paper, we present a deep multi-instance learning approach for person re-identification. Since most publicly available databases for pedestrian re-identification are not enough big, over-fitting problems occur in deep learning architectures. To tackle this problem, person re-identification is expressed as a deep multi-instance learning issue. Therefore, a multi-scale feature learning process is introduced which is driven by optimizing a novel cost function. We report on experiments and comparisons to other state-of-the-art algorithms using publicly available databases such as VIPeR and ETHZ.

Automatic Cartoon Colorization Based on Convolutional Neural Network

This paper deals with automatic cartoon colorization. This is a hard issue, since it is an ill-posed problem that usually requires user intervention to achieve high quality. Motivated by the recent successes in natural image colorization based on deep learning techniques, we investigate the colorization problem at the cartoon domain using Convolutional Neural Network. To our best knowledge, no existing papers or research studies address this problem using deep learning techniques. Here we investigate a deep Convolutional Neural Network based automatic color filling method for cartoons.

Detecting pedestrians in surveillance videos based on convolutional neural network and motion

Pedestrian detection is a fundamental computer vision task with many practical applications in robotics, video surveillance, autonomous driving, and automotive safety. However, it is still a challenging problem due to the tremendous variations in illumination, clothing, color, scale, and pose. The aim of this paper to present our dynamic pedestrian detector. In this paper, we propose a pedestrian detection approach that uses convolutional neural network (CNN) to differentiate pedestrian and non-pedestrian motion patterns. Although the CNN has good generalization performance, the CNN classifier is time-consuming. Therefore, we propose a novel architecture to reduce the time of feature extraction and training. Occlusion handling is one of the most important problem in pedestrian detection. For occlusion handling, we propose a method, which consists of extensive part detectors. The main advantage of our algorithm is that it can be trained on weakly labeled data, i.e. it does not require part annotations in the pedestrian bounding boxes.