Youngjung Uh

Service-oriented architecture based on biometric using random features and incremental neural networks

We propose a service-oriented architecture based on biometric system where training and classification tasks are used by millions of users via internet connection. Such a large-scale biometric system needs to consider template protection, accuracy and efficiency issues. This is a challenging problem since there are tradeoffs among these three issues. In order to simultaneously handle these issues, we extract both global and local features via controlling the sparsity of random bases without training. Subsequently, the extracted features are fused with a sequential classifier. In the proposed system, the random basis features are not stored for security reason. The non-training based on feature extraction followed by a sequential learning contributes to computational efficiency. The overall accuracy is consequently improved via an ensemble of classifiers. We evaluate the performance of the proposed system using equal error rate under a stolen-token scenario. Our experimental results show that the proposed method is robust over severe local deformation with efficient computation for simultaneous transactions. Although we focus on face biometrics in this paper, the proposed method is generic and can be applied to other biometric traits.

Efficient Multiview Stereo by Random-Search and Propagation

We present an efficient multi-view 3D reconstruction method based on randomization and propagation scheme. Our method progressively refines 3D point estimates by randomly perturbing the initial guess of 3D points and propagates photo-consistent ones to their neighbors. In contrast to previous refinement methods that perform local optimization for a better photo-consistency, our randomization approach takes lucky matchings for reducing the computational complexity. Experiments show favorable efficiency of the proposed method with the accuracy that is close to the state-of-the-art methods.

Generating panorama image by synthesizing multiple homography

This paper presents a method to generate image mosaics of a panoramic scene. In general, the relation between images which is required for mosaicing cannot be expressed by a single homography due to geometrical condition of the scene, even if the images are taken at the same position. Many existing methods are using only one homography to make panorama image while ignoring the geometrical variations. Therefore, they experience a lot of distortions and misalignments from input images which contain several planes which cannot be handled by one homography. In this paper, we present a novel method that utilizes synthesis of multiple homography to warp the images. Moreover, our method determines the number of homography automatically, without users input. By our method, various distortions of shapes and mismatches can be reduced.

Detection of Abnormal Behavior by Scene Analysis in Surveillance Video

In intelligent surveillance system, various methods for detecting abnormal behavior were proposed recently. However, most researches are not robust enough to be utilized for actual reality which often has occlusions because of assumption the researches have that individual objects can be tracked. This paper presents a novel method to detect abnormal behavior by analysing major motion of the scene for complex environment in which object tracking cannot work. First, we generate Visual Word and Visual Document from motion information extracted from input video and process them through LDA(Latent Dirichlet Allocation) algorithm which is one of document analysis technique to obtain major motion information(location, magnitude, direction, distribution) of the scene. Using acquired information, we compare similarity between motion appeared in input video and analysed major motion in order to detect motions which does not match to major motions as abnormal behavior.

Discovering overlooked objects

Only small additional training is required, except for baseline detector.The method improves detection accuracy by novel co-occurrence re-scoring.The method can detect new objects which were not to be found by other methods. Contextual detection not only uses visual features, but also leverages contextual information from the scene in the image. Most conventional context based methods have heavy training cost or large dependence on the original baseline detector. To overcome such shortcomings, we propose a new method based on co-occurrence context. It is built upon recent off-the-shelf baseline detector and achieves higher accuracy than existing works while detecting additional true positives which the baseline detector could not find. Furthermore we construct an indoor specific NYUv2-context dataset to investigate context-based detection of indoor objects. It is a subset of original NYU-depth-v2 dataset and to be published online to encourage context researches. In the experiment, the proposed method obtained 21.22% mAP which outperforms the baseline and compared context-based work by 0.91 and 0.36 percentage point mAP respectively.

Color and shape feature-based detection of speed sign in real-time

This paper presents a method for detecting speed sign based on color and shape features in real-time under real-life environment. In our method, Region Of Interest(ROI) is extracted and verified based on shape feature. In the first step, ROI is roughly extracted by segmentation of a red rim and the segments are optimized by the boundary using guided image filtering. Next step, the shape-based detection verifies the extracted red rim. We compare three different shape-based detection methods, RSD, BCT, and STVUE, and the RSD shows the best speed sign detection rate of 93% on the experimental data of 62 images containing 85 speed sign.

Real-time background subtraction based on GPGPU for high-resolution video surveillance

Demand for intelligent surveillance has been increasing, to automatically detect and prevent dangerous situations with surveillance cameras. Image analysis, the most essential element in intelligent surveillance system, has continuously developed and contributed to the improvement. To analyze surveillance videos, foreground segmentation is vital which require background modeling. This paper proposes background modeling method which is robust to illumination variation and shadow area. Also, the proposed method is applicable to high-resolution videos in real time with modification for GPU implementation. We validate our method on different types of dataset including our new benchmark dataset to analyze the result quantitatively and qualitatively. The execution time of proposed method is 228.2 FPS for High Definition videos with NVIDIA GTX660.

Weighing classes and streams: toward better methods for two-stream convolutional networks

Abstract. The emergence of two-stream convolutional networks has boosted the performance of action recognition by concurrently extracting appearance and motion features from videos. However, most existing approaches simply combine the features by averaging the prediction scores from each recognition stream without realizing that some classes favor greater weight for appearance than motion. We propose a fusion method of two-stream convolutional networks for action recognition by introducing objective functions of weights with two assumptions: (1) the scores from streams do not weigh the same and (2) the weights vary across different classes. We evaluate our method by extensive experiments on UCF101, HMDB51, and Hollywood2 datasets in the context of action recognition. The results show that the proposed approach outperforms the standard two-stream convolutional networks by a large margin (5.7%, 4.8%, and 3.6%) on UCF101, HMDB51, and Hollywood2 datasets, respectively.

Multi-view 3D reconstruction by random-search and propagation with view-dependent patch maps

This paper proposes an efficient multi-view 3D reconstruction method based on randomization and propagation scheme. Our method progressively refines a 3D model of a given scene by randomly perturbing the initial guess of 3D points and propagating photo-consistent ones to their neighbors. While finding local optima is an ordinary method for better photo-consistency, our randomization and propagation takes lucky matchings to spread better points replacing old ones for reducing the computational complexity. Experiments show favorable efficiency of the proposed method accompanied by competitive accuracy with the state-of-the-art methods.

Illumination invariant color segmentation method based on cluster center tree for traffic sign detection

This paper proposes a color segmentation method that can locate candidate regions of traffic signs accurately and reliably from real world images. In the real world, there are various light conditions which make the color segmentation very difficult problem. Hence, we propose an illumination invariant color segmentation method. The proposed method consists of two parts; 1) cluster center tree-based segmentation 2) illumination estimation. Cluster center tree is trained for color segmentation. Illumination estimation algorithm classifies light condition of the input images. We validate the proposed method qualitatively and quantitatively with 1,745 images containing red and blue traffic signs captured with four light conditions; sunny, cloudy, rainy and night. The proposed method achieves the high detection rate of 99.25% in sunny, 98.33% in cloudy, 87.85% in rainy and 88.70% at night.