Learning Semantic-Aware Local Features for Long Term Visual Localization.

Extracting robust and discriminative local features from images plays a vital role for long term visual localization, whose challenges are mainly caused by the severe appearance differences between matching images due to the day-night illuminations, seasonal changes, and human activities. Existing solutions resort to jointly learning both keypoints and their descriptors in an end-to-end manner, leveraged on large number of annotations of point correspondence which are harvested from the structure from motion and depth estimation algorithms. While these methods show improved performance over non-deep methods or those two-stage deep methods, i.e., detection and then description, they are still struggled to conquer the problems encountered in long term visual localization. Since the intrinsic semantics are invariant to the local appearance changes, this paper proposes to learn semantic-aware local features in order to improve robustness of local feature matching for long term localization. Based on a state of the art CNN architecture for local feature learning, i.e., ASLFeat, this paper leverages on the semantic information from an off-the-shelf semantic segmentation network to learn semantic-aware feature maps. The learned correspondence-aware feature descriptors and semantic features are then merged to form the final feature descriptors, for which the improved feature matching ability has been observed in experiments. In addition, the learned semantics embedded in the features can be further used to filter out noisy keypoints, leading to additional accuracy improvement and faster matching speed. Experiments on two popular long term visual localization benchmarks (Aachen Day and Night v1.1, Robotcar Seasons) and one challenging indoor benchmark (InLoc) demonstrate encouraging improvements of the localization accuracy over its counterpart and other competitive methods.

StATL2-like could affect growth and cold tolerance of plant by interacting with StCBFs.

KEY MESSAGE: Our results confirmed that StATL2-like could interact with StCBFs and regulate plant growth. Meanwhile, StATL2-like acted as a negative regulator on low-temperature tolerance in plants. As important transcription factors for resisting many kinds of stresses, C-repeat-binding factors (CBF) play a key role in plant low-temperature tolerance by increasing COR genes expressions. Here, we report that StATL2-like, a RING-H2 E3 ubiquitin in Solanum tuberosum L., interacted with StCBF1 and StCBF4, respectively. AtATL2 is a highly homologous gene of StATL2-like in Arabidopsis thaliana. Under normal conditions, atl2 Arabidopsis mutant showed a growth inhibition phenotype while overexpressed StATL2-like in wild type Arabidopsis and atl2 mutant promoted plant growth. Besides, atl2 mutant had better low-temperature tolerance compared with wild type and StATL2-like transgenic lines which demonstrated that StATL2-like acted as a negatively regulator on low-temperature tolerance in plant. Moreover, atl2 mutant improved the scavenging capacity of reactive oxygen species (ROS) and alleviate the damage of photosynthetic system II (PSII) compared with StATL2-like transgenic lines under cold conditions. These results suggested a new component in CBF-dependent pathway to regulate plant growth and response to low-temperature stress in potato plants.

Multiple Cayley-Klein metric learning.

As a specific kind of non-Euclidean metric lies in projective space, Cayley-Klein metric has been recently introduced in metric learning to deal with the complex data distributions in computer vision tasks. In this paper, we extend the original Cayley-Klein metric to the multiple Cayley-Klein metric, which is defined as a linear combination of several Cayley-Klein metrics. Since Cayley-Klein is a kind of non-linear metric, its combination could model the data space better, thus lead to an improved performance. We show how to learn a multiple Cayley-Klein metric by iterative optimization over single Cayley-Klein metric and their combination coefficients under the objective to maximize the performance on separating inter-class instances and gathering intra-class instances. Our experiments on several benchmarks are quite encouraging.

Exploring Local and Overall Ordinal Information for Robust Feature Description.

This paper aims to build robust feature descriptors by exploring intensity order information in a patch. To this end, the local intensity order pattern (LIOP) and the overall intensity order pattern (OIOP) are proposed to effectively encode intensity order information of each pixel in different aspects. Specifically, LIOP captures the local ordinal information by using the intensity relationships among all the neighbouring sampling points around a pixel, while OIOP exploits the coarsely quantized overall intensity order of these sampling points. These two kinds of patterns are then separately aggregated into different ordinal bins, leading to two kinds of feature descriptors. Furthermore, as these two kinds of descriptors could encode complementary ordinal information, they are combined together to obtain a discriminative and compact mixed intensity order pattern descriptor. All these descriptors are constructed on the basis of relative relationships of intensities in a rotationally invariant way, making them be inherently invariant to image rotation and any monotonic intensity changes. Experimental results on image matching and object recognition are encouraging, demonstrating the superiorities of our descriptors over the state of the art.

Algebraic Error Based Triangulation and Metric of Lines.

Line triangulation, a classical geometric problem in computer vision, is to determine the 3D coordinates of a line based on its 2D image projections from more than two views of cameras with known projection matrices. Compared to point features, line segments are more robust to matching errors, occlusions, and image uncertainties. In addition to line triangulation, a better metric is needed to evaluate 3D errors of line triangulation. In this paper, the line triangulation problem is investigated by using the Lagrange multipliers theory. The main contributions include: (i) Based on the Lagrange multipliers theory, a formula to compute the Plücker correction is provided, and from the formula, a new linear algorithm, LINa, is proposed for line triangulation; (ii) two optimal algorithms, OPTa-I and OPTa-II, are proposed by minimizing the algebraic error; and (iii) two metrics on 3D line space, the orthogonal metric and the quasi-Riemannian metric, are introduced for the evaluation of line triangulations. Extensive experiments on synthetic data and real images are carried out to validate and demonstrate the effectiveness of the proposed algorithms.

[The backgroud sky subtraction around [OIII] line in LAMOST QSO spectra].

At present, most sky-subtraction methods focus on the full spectrum, not the particular location, especially for the backgroud sky around [OIII] line which is very important to low redshift quasars. A new method to precisely subtract sky lines in local region is proposed in the present paper, which sloves the problem that the width of Hß-[OIII] line is effected by the backgroud sky subtraction. The exprimental results show that, for different redshift quasars, the spectral quality has been significantly improved using our method relative to the original batch program by LAMOST. It provides a complementary solution for the small part of LAMOST spectra which are not well handled by LAMOST 2D pipeline. Meanwhile, This method has been used in searching for candidates of double-peaked Active Galactic Nuclei.

Weighted similarity-invariant linear algorithm for camera calibration with rotating 1D objects.

In this paper, a weighted similarity-invariant linear algorithm for camera calibration with rotating 1D objects is proposed. First, we propose a new estimation method for computing the relative depth of the free endpoint on the 1D object and prove its robustness against noise compared with those used in previous literature. The introduced estimator is invariant to image similarity transforms, resulting in a similarity-invariant linear calibration algorithm which is slightly more accurate than the well-known normalized linear algorithm. Then, we use the reciprocals of the standard deviations of the estimated relative depths from different images as the weights on the constraint equations of the similarity-invariant linear calibration algorithm, and propose a weighted similarity-invariant linear calibration algorithm with higher accuracy. Experimental results on synthetic data as well as on real image data show the effectiveness of our proposed algorithm.

Rotationally invariant descriptors using intensity order pooling.

This paper proposes a novel method for interest region description which pools local features based on their intensity orders in multiple support regions. Pooling by intensity orders is not only invariant to rotation and monotonic intensity changes, but also encodes ordinal information into a descriptor. Two kinds of local features are used in this paper, one based on gradients and the other on intensities; hence, two descriptors are obtained: the Multisupport Region Order-Based Gradient Histogram (MROGH) and the Multisupport Region Rotation and Intensity Monotonic Invariant Descriptor (MRRID). Thanks to the intensity order pooling scheme, the two descriptors are rotation invariant without estimating a reference orientation, which appears to be a major error source for most of the existing methods, such as Scale Invariant Feature Transform (SIFT), SURF, and DAISY. Promising experimental results on image matching and object recognition demonstrate the effectiveness of the proposed descriptors compared to state-of-the-art descriptors.

[Galaxy spectrum subtraction of a mixed spectrum based on two class PCA eigen-spectra].

The authors present a new method called two class PCA for decomposing the mixed spectra, namely, for subtracting the host galaxy contamination from each SN spectrum. The authors improved the quality of reconstructed galaxy spectrum and computational efficiency, and these improvements were realized because we used both the PCA eigen spectra of galaxy templates library and SN templates library to model the mixed spectrum. The method includes mainly three steps described as follows. The first step is calculating two class PCA eigen spectra of galaxy templates and SN templates respectively. The second step is determining all reconstructed coefficients by the SVD matrix decomposition or orthogonal transformation. And the third step is computing a reconstructed galaxy spectrum and subtracting it from each mixed spectrum. Experiments show that this method can obtain an accurate decomposition of a mixed synthetic spectrum, and is a method with low time-consumption to get the reliable SN spectrum without galaxy contamination and can be used for spectral analysis of large amount of spectra. The time consumption using our method is much lower than that using chi2-template fitting for a spectrum.

[Kernel regression application in estimating stellar fundamental parameters].

The three fundamental parameters of stellar atmosphere, i.e. the effective temperature, the surface gravity, and the metallic, determine the continuum and spectral lines in the stellar spectrum. With the development of the modern telescopes such as SDSS, LAMOST projects, the great voluminous spectra demand to explore automatic celestial spectral analysis methods. It is most significant for Galaxy research to develop automatic methods determining the fundamental parameters from stellar spectra data. Two non-linear regression algorithms, kernel least squared regression (KLSR) and kernel PCA regression (KPCR), are proposed for estimating the three parameters in the present paper. The linear regression models, LSR and PCR, are extended to non-linear regression by using a kernel function for the stellar parameter estimation from spectra. Extensive experiments on low resolution spectra data show: (1) KLSR and KPCR methods realize the regression from spectrum to the effective temperature and gravity. KLSR is sensitive to the noise while KPCR is robust than the former. (2) For the effective temperature estimation, the two algorithms perform similarly; and for the gravity and metallic estimation, the KPCR is superior to the KLSR and the NPR (Non-parameter regression); (3) KLSR and KPCR methods are simple and efficient for the stellar spectral parameter estimation.

[Reducing the searching range of supernova candidates automatically in a flood of spectra of galaxies].

Supernova (SN) is one of the most intense astronomical phenomena among the known stellar activities, but compared with several billion astronomical objects which people have probed, the number of supernova the authors have observed is very small. Therefore, the authors need to find faster and higher-efficiency approaches to searching supernova. In the present paper, we present a novel automated method, which can be successfully used to reduce the range of searching for 1a supernova candidates in a huge number of galaxy spectra. The theoretical basis of the method is clustering and outlier picking, by introducing and measuring local outlier factors of data samples, description of statistic characters of SN emerges in low dimension space. Firstly, eigenvectors of Peter's 1a supernova templates are acquired through PCA projection, and the description of la supernova's statistic characters is calculated. Secondly, in all data set, the local outlier factor (LOF) of each galaxy is calculated including those SN and their host galaxy spectra, and all LOFs are arranged in descending order. Finally, spectra with the largest first one percent of all LOFs should be the reduced 1a SN candidates. Experiments show that this method is a robust and correct range reducing method, which can get rid of the galaxy spectra without supernova component automatically in a flood of galaxy spectra. It is a highly efficient approach to getting the reliable candidates in a spectroscopy survey for follow-up photometric observation.

[The application of adaptive boosting method in automated spectral classification of active galactic nuclei].

Given a set of low-redshift spectra of active galactic nuclei, the wave bands of spectra in the rest frame were intercepted according to the different features of emission lines of broad-line AGNs and narrow-line AGNs, and an adaptive boosting (Adaboost) method was developed to carry out the classification experiments of feature fusion. As a result, the wave band of Halpha and [N II] was confirmed to be the main discriminative feature between broad-line AGNs and narrow-line AGNs. Then based on the wave band of Halpha and [N II], the Adaboost method was used for the spectral classification. In this method, the "weak classifiers" were increased constantly during training until a scheduled error rate or a maximum cycle times was met, then the classification judgment of the consequent collective classifier was determined by the votes of respective judgments of these "weak classifiers". The Adaboost method needs not to adjust parameters in advance and the results of "weak classifiers" are only required to be better than random guessing, so its algorithm is very simple. As proved by the experiments, the adaboost method achieves good performance in the classification just based on the wave band of Halpha and [N II] so that it could be applied effectively to the automatic classification of large amount of AGN spectra from the large-scale spetral surveys.

[Celestial spectrum flux standardization for classification].

Celestial spectra should be preprocessed before automated classification to eliminate the disturbance of noise, observa-tion environment, and flux aberrance. In the present work, the authors studied the spectrum flux standardization problem. By analyzing the disturbing factors and their characteristics, the authors put forward a theoretical model for spectra flux, and corre-spondingly give several flux standardizing methods. The rationality/correctness of the model, and the satisfactory performance of the proposed methods have been obtained by the experiments over normal galaxies (NGs) and quasi-stellar object (Qso). Furthermore, the authors theoretically analyze, compare and evaluate them. In particular, this work indicated that the conventional method is worse than the proposed one. And the investigation is also particularly significant for other automatic spectrum processing study, e. g. redshift determination, effective temperature, metallic estimation, etc.

[Studies of spectra classification based on kernel covering algorithm].

A kernel based covering algorithm, called the kernel covering algorithm (KCA), is proposed for the classification of celestial spectra. This algorithm is a combination of kernel trick with the covering algorithm, and is used to extract the support vectors in feature space. The experiments show that the classification result based on KCA is a little less than that based on SVM. However, KCA only involves the distance computation without the need to solve the quadratic programming problem. Also, KCA is insensitive to the width of gauss window. Although KCA has a comparable classification performance with the covering algorithm, it changes the distance between samples in feature space by the nonlinear mapping such that the distribution of samples is more adaptable to classify. Therefore, the number of KCA's resulting support vectors is significantly smaller than that of the covering algorithm.

[A method for obtaining redshifts of quasars based on wavelet multi-scaling feature matching].

The LAMOST project, the world's largest sky survey project being implemented in China, is expected to obtain 10(5) quasar spectra. The main objective of the present article is to explore methods that can be used to estimate the redshifts of quasar spectra from LAMOST. Firstly, the features of the broad emission lines are extracted from the quasar spectra to overcome the disadvantage of low signal-to-noise ratio. Then the redshifts of quasar spectra can be estimated by using the multi-scaling feature matching. The experiment with the 15, 715 quasars from the SDSS DR2 shows that the correct rate of redshift estimated by the method is 95.13% within an error range of 0. 02. This method was designed to obtain the redshifts of quasar spectra with relative flux and a low signal-to-noise ratio, which is applicable to the LAMOST data and helps to study quasars and the large-scale structure of the universe etc.

[A novel method for the determination of redshifts of normal galaxies by non-linear dimensionality reduction].

It is difficult to determine the redshifts of normal galaxies (NG) from their spectra because of their common weak absorption property. In the present work, a novel method is proposed to effectively deal with this issue. The proposed method is composed of the following three parts: At first, the wavelet transform coefficients at the fourth scaling are experimentally found to be appropriate and used as our features to represent the absorption information from NG absorption lines, break points, and absorption bands. Then, the features are mapped by a non-linear method, LLE (locally linear embedding), onto an one-dimensional manifold in the 3D space; Finally, the NG redshifts are obtained by the nearest neighborhood technique from the redshift distribution on the manifold. Besides, the proposed method is compared with widely used PCA method in the literature with SDSS database, and is shown to be more accurate for the redshifts determination.

[Automated recognition of quasars based on adaptive radial basis function neural networks].

Recognizing and certifying quasars through the research on spectra is an important method in the field of astronomy. This paper presents a novel adaptive method for the automated recognition of quasars based on the radial basis function neural networks (RBFN). The proposed method is composed of the following three parts: (1) The feature space is reduced by the PCA (the principal component analysis) on the normalized input spectra; (2) An adaptive RBFN is constructed and trained in this reduced space. At first, the K-means clustering is used for the initialization, then based on the sum of squares errors and a gradient descent optimization technique, the number of neurons in the hidden layer is adaptively increased to improve the recognition performance; (3) The quasar spectra recognition is effectively carried out by the above trained RBFN. The author's proposed adaptive RBFN is shown to be able to not only overcome the difficulty of selecting the number of neurons in hidden layer of the traditional RBFN algorithm, but also increase the stability and accuracy of recognition of quasars. Besides, the proposed method is particularly useful for automatic voluminous spectra processing produced from a large-scale sky survey project, such as our LAMOST, due to its efficiency.

[A method for auto-extraction of spectral lines based on convolution type of wavelet packet transformation].

The important astrophysical information is hidden in spectral lines of astronomical spectra. The presen paper presents a method for auto-extraction of spectral lines based on convolution type of wavelet packet. This method consists of four main steps: First, the observed spectra are transformed by convolution type of wavelet packet with 4th scale. Then, the noise with coefficients of the 4th scale is eliminated by the local correlation algorithm and threshold in the wavelet packet domain. After that, middle and high frequency coefficients are selected to reconstruct the feature of the spectral lines. Finally, with the reconstructed feature of the spectral lines, spectral lines in observed spectra are searched. The results of our experiments, which include the spectral lines of stars, normal galaxies and active galaxies, show that the method can robustly and accurately extract the spectral lines. The method was applied to extract the SDSS spectral lines and compute the redshifts with those lines. By comparing the redshifts with those given by SDSS, the extraction has proven successful and practical.

[Spectra classification based on generalized discriminant analysis].

A kernel based generalized discriminant analysis (GDA) technique is proposed for the classification of stars, galaxies, and quasars. GDA combines the LDA algorithm with kernel trick, and samples are projected by nonlinear mapping onto the feature space F with high dimensions, and then LDA is conducted in F. Also, it could be inferred that GDA which combines the extension of Fisher's criterion with kernel trick is complementary to kernel Fisher discriminant framework. LDA, GDA, PCA and KPCA were experimentally compared with these three different kinds of spectra. Among these four techniques, GDA obtains the best result, followed by LDA, and PCA is the worst. Although KPCA is also a kernel based technique, its performance is not satisfactory if the selected number of the principal components is small, and in some cases, it appears even worse than LDA, a non-kernel based technique.

[Using neural networks based template matching method to obtain redshifts of normal galaxies].

Galaxies can be divided into two classes: normal galaxy (NG) and active galaxy (AG). In order to determine NG redshifts, an automatic effective method is proposed in this paper, which consists of the following three main steps: (1) From the template of normal galaxy, the two sets of samples are simulated, one with the redshift of 0.0-0.3, the other of 0.3-0.5, then the PCA is used to extract the main components, and train samples are projected to the main component subspace to obtain characteristic spectra. (2) The characteristic spectra are used to train a Probabilistic Neural Network to obtain a Bayes classifier. (3) An unknown real NG spectrum is first inputted to this Bayes classifier to determine the possible range of redshift, then the template matching is invoked to locate the redshift value within the estimated range. Compared with the traditional template matching technique with an unconstrained range, our proposed method not only halves the computational load, but also increases the estimation accuracy. As a result, the proposed method is particularly useful for automatic spectrum processing produced from a large-scale sky survey project.