Computer Vision - ACCV 2012 Workshops

Gradient calculation and edge detection are well-known problems in image processing and the fundament for many approaches for line detection, segmentation, contour extraction, or model fitting. A large variety of algorithms for edge detection already exists but strong image noise is still a challenge. Especially in automatic surveillance and reconnaissance applications with visual-optical, infrared, or SAR imagery, high distance to objects and weak signal-to-noise-ratio are difficult tasks to handle. In this paper, a new approach using Local Binary Patterns (LBPs) is presented, which is a crossover between texture analysis and edge detection. It shows similar results as the Canny edge detector under normal conditions but performs better in presence of noise. This characteristic is evaluated quantitatively with different artificially generated types and levels of noise in synthetic and natural images.

Local binary patterns [LBP][1] are popular texture descriptors in many image analysis tasks. One of the important aspects of this texture descriptor is their rotational invariance. Most work in LBP has focused on 2D images. Here, we present a three dimensional LBP with a rotational invariant operator using spherical harmonics. Unlike Fehr and Burkhardt [2], the invariance is constructed implicitly, without considering all possible combinations of the pattern. We demonstrate the 3D LBP on phantom data and a clinical CTA dataset.

Dynamic textures are image sequences recording texture in motion. Given a sample video, the goal of synthesis is to create a new sequence enlarged in spatial and/or temporal domain, which looks perceptually similar to the input. Most synthesis methods are mainly focused on extending sequences only in the temporal domain. In this paper, we propose a dynamic texture synthesis approach for spatial domain, where we aim to enlarge the frame size while preserving the aspect and motion of the original video. For this purpose, we use a patch-based synthesis method based on LBP-TOP features. In our approach, 3D patch regions from the input are selected and copied to an output sequence. Usually, in other patch-based approaches, the selection of the patches is based only in the color, which cannot capture the spatial and temporal information, causing an unnatural look in the output. In contrast, we propose to use the LBP-TOP operator, which implicitly represents information about appearance, dynamics and correlation between frames. The experiments show that the use of the LBP-TOP improves the performance of other methods giving a good description of the structure and motion of dynamic textures without generating visible discontinuities or artifacts.

The correntropy is originally proposed to measure the similarity between two random variables and developed as a novel metrics for feature matching. As a kernel method, the parameter of kernel function is very important for correntropy metrics. In this paper, we propose an adaptive parameter selection strategy for correntropy metrics and deduce a close-form solution based on the Maximum Correntropy Criterion (MCC). Moreover, considering the correlation of localized features, we modify the classic correntropy into a block-wise metrics. We verify the proposed metrics in face recognition applications taking Local Binary Pattern (LBP) features. Combined with the proposed adaptive parameter selection strategy, the modified block-wise correntropy metrics could result in much better performance in the experiments.

A new method to describe sperm vitality using a hybrid combination of local and global texture descriptors is proposed in this paper. In this regard, a new adaptive local binary pattern (ALBP) descriptor is presented in order to carry out the local description. It is built by adding oriented standard deviation information to an ALBP descriptor in order to achieve a more complete representation of the images and hence it has been called ALBPS. Regarding semen vitality assessment, ALBPS outperformed previous literature works with an 81.88% of accuracy and it also yielded higher hit rates than the LBP and ALBP base-line methods. Concerning the global description of sperm heads, several classical texture algorithms were tested and a descriptor based on Wavelet transform and Haralick feature extraction (WCF13) obtained the best results. Both local and global descriptors were combined and the classification was carried out with a Support Vector Machine. Therefore, our proposal is novel in three ways. First, a new local feature extraction method ALBPS is introduced. Second, a hybrid method combining the proposed local ALBPS and a global descriptor is presented outperforming our first approach and all other methods evaluated for this problem. Third, vitality classification accuracy is greatly improved with the two former texture descriptors presented. F-Score and accuracy values were computed in order to measure the performance. The best overall result was yielded by combining ALBPS with WCF13 reaching a F-Score equals to 0.886 and an accuracy of 85.63%.

Background subtraction is the first step in many video surveillance systems, its performance has a decisive influence on the result of the post-processing. An effective background subtraction algorithm should distinguish foreground from the background sensitively, and adapt to the variation of background scenes robustly, such as illumination changes or dynamic scenes. In this paper, a novel pixel-wise background subtraction algorithm is introduced. First, we propose a novel texture descriptor named Multi-Channel Scale Invariant Local Ternary Pattern(MC-SILTP). The pattern is cross-calculated in RGB color channels with the Scale Invariant Local Ternary Pattern operator. This descriptor does not only show an excellent performance in abundant texture regions, but also in flat regions. Secondly, we model each background pixel with a codebook rather than estimating the probability density functions. The codebook is consisted of many MC-SILTP samples actually observed in the past. A lot of experiments have been done over the proposed approach, results indicates that this approach is well balanced in sensitivity and robustness. It can handle the tricky problem of illumination changes robustly while detecting complete objects in flat areas sensitively. Comparison between the proposed one and several popular background subtraction algorithms demonstrates that it outperforms the state-of-the-art.

In this paper, we propose a novel variant of Local Binary Patterns (LBP) so-called Elliptical Local Binary Patterns (ELBP) for face recognition. In ELBP, we use horizontal and vertical ellipse patterns to capture micro facial feature of face images in both horizontal and vertical directions. ELBP is applied in face recognition with dimension reduction step by Whitened Principal Component Analysis (WPCA). Our experiment results upon AR, FERET and Surveillance Cameras Face (SCface) databases prove the advantages of ELBP over LBP for face recognition under different conditions and with ELBP WPCA we can get very remarkable results.

A local matching approach, known as Electoral College, where each block contributes one single vote to the final decision, which is generated by a simply majority voting from all local binary decisions, has been proved to be stable for political elections as well as general pattern recognition. Given the registration difficulties caused by the non-rigidity of human face images, block LBP displacement is introduced so that an Electoral College, where a local decision is made on LBP statistics for each block, can be applied to face recognition problems. Extensive experiments are carried out and have demonstrated the outstanding performances of the block LBP displacement based Electoral College in comparison with the original LBP approach. It is expected and shown by experiments that the approach also applies to descriptor approaches other than LBP.

In this paper, we propose Learned Local Curvelet Patterns (LLCP) for presenting the local features of facial images. The proposed method is based on curvelet transform which can overcome the weakness of traditional Gabor wavelets in higher dimension, and better capture the curve singularities and hyperplane singularities of facial images. Different from wavelet transform, curvelet transform can effectively and efficiently approximate the curved edges with very few coefficients as well as taking space-frequency information into consideration. First, LLCP designs several learned codebooks from Curvelet filtered facial images. Then each facial image can be encoded into multiple pattern maps and finally block-based histograms of these patterns are concatenated into an histogram sequence to be used as a face descriptor. In order to reduce the face feature descriptor, 2-Directional L1-Norm Based 2DPCA ((2D)²PCA-L1) is proposed which is simultaneously considering the row and column directions for efficient face representation and recognition. Performance assessment in several face recognition problem shows that the proposed approach is superior to traditional ones.

User authentication is an important step to protect information and in this field face biometrics is advantageous. Face biometrics is natural, easy to use and less human-invasive. Unfortunately, recent work has revealed that face biometrics is vulnerable to spoofing attacks using low-tech cheap equipments. This article presents a countermeasure against such attacks based on the LBP − TOP operator combining both space and time information into a single multiresolution texture descriptor. Experiments carried out with the REPLAY ATTACK database show a Half Total Error Rate (HTER) improvement from 15.16% to 7.60%.

RGB-D is a powerful source of data providing the aligned depth information which has great potentials in improving the performance of various problems in image understanding, while Local Binary Patterns (LBP) have shown excellent results in representing faces. In this paper, we propose a novel efficient LBP-based descriptor, namely Gradient-LBP (G-LBP), specialized to encode the facial depth information inspired by 3DLBP, yet resolves its inherent drawbacks. The proposed descriptor is applied to gender recognition task and shows its superiority to 3DLBP in all the experimental setups on both Kinect and range scanner databases. Furthermore, a weighted combination scheme of the proposed descriptor for depth images and the state-of-the-art LBP ^U2 for grayscale images applied in gender recognition is proposed and evaluated. The result reinforces the effectiveness of the proposed descriptor in complementing the source of information from the luminous intensity. All the experiments are carried out on both the high quality 3D range scanner database - Texas 3DFR and images of lower quality obtained from Kinect - EURECOM Kinect Face Dataset to show the consistency of the performance on different sources of RGB-D data.

While there is a significant number of works addressing e.g. pose and illumination variation problems in face recognition, the vulnerabilities to spoofing attacks were mostly unexplored until very recently when an increasing attention is started to be paid to this threat. A spoofing attack occurs when a person tries to masquerade as someone else e.g. by wearing a mask to gain illegitimate access and advantages. This work provides the first investigation in research literature on the use of dynamic texture for face spoofing detection. Unlike masks and 3D head models, real faces are indeed non-rigid objects with contractions of facial muscles which result in temporally deformed facial features such as eye lids and lips. Our key idea is to learn the structure and the dynamics of the facial micro-textures that characterise only real faces but not fake ones. Hence, we introduce a novel and appealing approach to face spoofing detection using the spatiotemporal (dynamic texture) extensions of the highly popular local binary pattern approach. We experiment with two publicly available databases consisting of several fake face attacks of different natures under varying conditions and imaging qualities. The experiments show excellent results beyond the state-of-the-art.

This paper proposes a class-specified segmentation method, which can not only segment foreground objects from background at pixel level, but also parse images. Such class-specified segmentation is very helpful to many other computer vision tasks including computational photography. The novelty of our method is that we use multi-scale superpixels to effectively extract object-level regions instead of using only single scale superpixels. The contextual information across scales and the spatial coherency of neighboring superpixels in the same scale are represented and integrated via a Conditional Random Field model on multi-scale superpixels. Compared with the other methods that have ever used multi-scale superpixel extraction together with across-scale contextual information modeling, our method not only has fewer free parameters but also is simpler and effective. The superiority of our method, compared with related approaches, is demonstrated on the two widely used datasets of Graz02 and MSRC.

Auto white-balance plays a very important role in computer vision, and also is a prerequisite of color processing algorithms. For keeping the color constancy in the real-time outdoor environment, a simple and flexible auto white balance algorithm based on the color histogram overlap of the image is presented in this paper. After looking at a numerous images under different illuminance, an essential characteristic of the white-balance, the color histogram coincidence, is generalized as the basic criterion. Furthermore the overlap area of the color histogram directly reflects this coincidence, namely, when the overlap area of the color histogram reaches the maximum, the respective gain coefficients of color channels can be derived to achieve the white-balance of the camera. Through the subjective and objective evaluations based on the processing of real world images, the proposed histogram overlap algorithm can not only flexibly implement the auto white-balance of the camera but also achieve the outstanding performance in the real-time outdoor condition.

We have developed a robust statistical edge detection method by combining the ideas of Kundus method, in which the region segmentation of local area is used, and Fukuis method, in which a statistic evaluation value separability is used for edge extraction and also have developed a region segmentation method based on the global features like the statistics of the region. A new region segmentation method is developed by combining these two methods, in which the edge extraction method is used instead of the first step of region segmentation method. We obtained the almost same results as the ones of previous region segmentation method. The proposed one has some advantages because we are able to introduce a new conspicuity degree including a clear contrast value with the adjacent regions, a envelopment degree based on clear edge and so on without much difficulty and it will contribute to develop a further unification algorithm and a new feature extraction method for scene recognition.

In this paper we describe a close-to-sensor low latency visual processing system. We show that by adaptively sampling visual information, low level tracking can be achieved at high temporal frequencies with no increase in bandwidth and using very little memory. By having close-to-sensor processing, image regions can be captured and processed at millisecond sub-frame rates. If spatiotemporal regions have little useful information in them they can be discarded without further processing. Spatiotemporal regions that contain ‘interesting’ changes are further processed to determine what the interesting changes are. Close-to-sensor processing enables low latency programming of the image sensor such that interesting parts of a scene are sampled more often than less interesting parts. Using a small set of low level rules to define what is interesting, early visual processing proceeds autonomously. We demonstrate system performance with two applications. Firstly, to test the absolute performance of the system, we show low level visual tracking at millisecond rates and secondly a more general recursive Baysian tracker.

Colorimetric correction is a necessary task to generate comfortable stereoscopic images. This correction is usually performed with a 3D lookup table that can correct images in real-time and can deal with the non-independence of the colour channels. In this paper, we present a method to compute such 3D lookup table with a non-linear process that minimizes the colorimetric properties of the images. This lookup table is represented by a polynomial basis to reduce the number of required parameters. We also describe some optimizations to speedup the processing time.

In this paper we present an easy method for multiple camera calibration with common field of view only from vertical lines. The locations of the vertical lines are known in advance. Compared to other calibration objects, the vertical lines have some good properties, since they can be easily built and can be visible by cameras in any direction simultaneously. Given 5 fixed vertical lines, an image containing them taken by a camera may provide 2 constraints in the intrinsic parameters of the camera, and extrinsic parameters can then be recovered. The calibration procedure consists of three main steps: Firstly, the image is rectified by a homography, which makes the projections of vertical lines parallel to u-axis in the rectified image. Secondly, for any vertical scan line in the rectified image, if we consider the scan line is taken by a virtual 1D camera, then we can calibrate the 1D camera. Finally, the intrinsic parameters of the original camera can be determined from the intrinsic parameters of the virtual 1D camera. By evaluating on both simulated and real data we demonstrate that our method is efficient and robust.

This paper proposes a method for long-range vehicle localization using fusion of omnidirectional camera and Global Positioning System (GPS) in wide urban environments. The main contributions are twofold: first, the positions estimated by visual sensor overcome the motion blur effects. The motion constrains of successive frames are obtained accurately under various scene structures and conditions. Second, the cumulative errors of visual odometry system are solved completely based on the fusion of local (visual odometry) and global positioning system. The visual odometry can yield the correct local position at short distance of movements but it will accumulate errors overtime, on the contrary, the GPS can yields the correct global positions but the local positions may be drifted. Moreover, the signals received from satellites are affected by multi-path and forward diffraction then the position errors increase when vehicles move in dense building regions or jump/miss in tunnels. To utilize the advantages of two sensors, the position information should be evaluated before fusion by Extended Kalman Filter (EKF) framework. This multiple sensor system can also compensate each other in the case of losing one of two. The simulation results demonstrate the accuracy of vehicle positions in long-range movements.

Development processes for building image recognition systems are highly specialized and require expensive expert knowledge. Despite some effort in developing generic image recognition systems, use of computer vision technology is still restricted to experts. We propose a flexible image recognition system (FOREST), which requires no prior knowledge about the recognition task and allows non-expert users to build custom image recognition systems, which solve a specific recognition task defined by the user. It provides a simple-to-use graphical interface which guides users through a simple development process for building a custom recognition system. FOREST integrates a variety of feature descriptors which are combined in a classifier using a boosting approach to provide a flexible and adaptable recognition framework. The evaluation shows, that image recognition systems developed with this framework are capable of achieving high recognition rates.

The majority of vision research focusses on advancing technical methods for image analysis, with a coupled increase in complexity and sophistication. The problem of providing access to these sophisticated techniques is largely ignored, leading to a lack of application by mainstream applications. We present a feature-based clustering segmentation algorithm with novel modifications to fit a developer-centred abstraction. This abstraction acts as an interface which accepts a description of segmentation in terms of properties (colour, intensity, texture, etc.), constraints (size, quantity) and priorities (biasing a segmentation). This paper discusses the modifications needed to fit the algorithm into the abstraction, which conditions of the abstraction it supports, and results of the various conditions demonstrating the coverage of the segmentation problem space. The algorithm modification process is discussed generally to help other researchers mould their algorithms to similar abstractions.

The integral histogram for images is an efficient preprocessing method for speeding up diverse computer vision algorithms including object detection, appearance-based tracking, recognition and segmentation. Our proposed Graphics Processing Unit (GPU) implementation uses parallel prefix sums on row and column histograms in a cross-weave scan with high GPU utilization and communication-aware data transfer between CPU and GPU memories. Two different data structures and communication models were evaluated. A 3-D array to store binned histograms for each pixel and an equivalent linearized 1-D array, each with distinctive data movement patterns. Using the 3-D array with many kernel invocations and low workload per kernel was inefficient, highlighting the necessity for careful mapping of sequential algorithms onto the GPU. The reorganized 1-D array with a single data transfer to the GPU with high GPU utilization, was 60 times faster than the CPU version for a 1K ×1K image reaching 49 fr/sec and 21 times faster for 512×512 images reaching 194 fr/sec. The integral histogram module is applied as part of the likelihood of features tracking (LOFT) system for video object tracking using fusion of multiple cues.

It is easy to retrieve the small size parts from small videos. It is also easy to retrieve the middle size part from large videos. However, we have difficulties to retrieve the small size parts from large videos. We have large needs for estimating plays in sport videos. Plays in sports are described as the motions of players. This paper proposes the play retrieving method based on both of the motion compensation vector and normal color frames in MPEG sports videos. In MPEG videos, there are motion compensation vectors. Using the motion compensation vectors, we do not need to estimate the motion vectors between adjacent frames. This leads to decrease the huge computations about motion estimations. This work uses the 1-dimensional degenerated descriptions of each motion image between 2 adjacent frames. Connecting the 1-dimensional degenerated descriptions on time direction, we have the space-time map. This space-time map describes a sequence of frames as a 2-dimensional image. Using this space-time map on motion compensation vector frames and normal color frames, this work shows the method to retrieve a small number of plays in a huge number of frames. Our experiment records 0.93 as recall, 0.81 as precision and 0.86 as F-measure on 139 plays in 132503 frames.

Most of video-surveillance based applications use a foreground extraction algorithm to detect interest objects from videos provided by static cameras. This paper presents a benchmark dataset and evaluation process built from both synthetic and real videos, used in the BMC workshop (Background Models Challenge). This dataset focuses on outdoor situations with weather variations such as wind, sun or rain. Moreover, we propose some evaluation criteria and an associated free software to compute them from several challenging testing videos. The evaluation process has been applied for several state of the art algorithms like gaussian mixture models or codebooks.

Background models are often used in video surveillance systems to find moving objects in an image sequence from a static camera. These models are often built under the assumption that the foreground objects are not known in advance. This assumption has led us to model background using one-class SVM classifiers. Our model belongs to a family of block-based nonparametric models that can be used effectively for highly complex scenes of various background distributions with almost the same configuration parameters for all examined videos. Experimental results are reported on a variety of test videos from the Background Models Challenge (BMC) competition.

The Mixture of Gaussians (MoG) is a frequently used method for foreground-background separation. In this paper, we propose an on-line learning framework that allows the MoG algorithm to quickly adapt its localized parameters. Our main contributions are: local parameter adaptations, a feedback based updating method for stopped objects, and hierarchical SURF features matching based ghosts and local illumination suppression method. The proposed model is rigorously tested and compared with several previous models on BMC data set and has shown significant performance improvements.

Foreground detection is the first step in video surveillance system to detect moving objects. Robust Principal Components Analysis (RPCA) shows a nice framework to separate moving objects from the background. The background sequence is then modeled by a low rank subspace that can gradually change over time, while the moving foreground objects constitute the correlated sparse outliers. In this paper, we propose to use a low-rank matrix factorization with IRLS scheme (Iteratively reweighted least squares) and to address in the minimization process the spatial connexity and the temporal sparseness of moving objects (e.g. outliers). Experimental results on the BMC 2012 datasets show the pertinence of the proposed approach.

This paper presents a novel saliency detection method and apply it to motion detection. Detection of salient regions in videos or images can reduce the computation power which is needed for complicated tasks such as object recognition. It can also help us to preserve important information in tasks like video compression. Recent advances have given birth to biologically motivated approaches for saliency detection. We perform salience estimation by measuring the change in pixel’s intensity value within a temporal interval while performing a filtering step via principal component analysis that is intended to suppress noise. We applied the method to Background Models Challenge (BMC) video data set. Experiments show that the proposed method is apt and accurate. Additionally, the method is fast to compute.

We present a robust background model for object detection and report its evaluation results using the database of Background Models Challenge (BMC). Our background model is based on a statistical local feature. In particular, we use an illumination invariant local feature and describe its distribution by using a statistical framework. Thanks to the effectiveness of the local feature and the statistical framework, our method can adapt to both illumination and dynamic background changes. Experimental results, which are done thanks to the database of BMC, show that our method can detect foreground objects robustly against background changes.