Unsupervised depth learning takes the appearance difference between a target view and a view synthesized from its adjacent frame as supervisory signal. Since the supervisory signal only comes from images themselves, the resolution of training data significantly impacts the performance. High-resolution images contain more fine-grained details and provide more accurate supervisory signal. However, due to the limitation of memory and computation power, the original images are typically down-sampled during training, which suffers heavy loss of details and disparity accuracy. In order to fully explore the information contained in high-resolution data, we propose a simple yet effective dual networks architecture, which can directly take high-resolution images as input and generate high-resolution and high-accuracy depth map efficiently. We also propose a Self-assembled Attention (SA-Attention) module to handle low-texture region. The evaluation on the benchmark KITTI and Make3D datasets demonstrates that our method achieves state-of-the-art results in the monocular depth estimation task.

Recently unsupervised learning of depth from videos has made remarkable progress and the results are comparable to fully supervised methods in outdoor scenes like KITTI. However, there still exist great challenges when directly applying this technology in indoor environments, e.g., large areas of non-texture regions like white wall, more complex ego-motion of handheld camera, transparent glasses and shiny objects. To overcome these problems, we propose a new optical-flow based training paradigm which reduces the difficulty of unsupervised learning by providing a clearer training target and handles the non-texture regions. Our experimental evaluation demonstrates that the result of our method is comparable to fully supervised methods on the NYU Depth V2 benchmark. To the best of our knowledge, this is the first quantitative result of purely unsupervised learning method reported on indoor datasets.

In recent years, some biorthogonal Catmull-Clark subdivision wavelet transforms constructed via the lifting scheme have been proposed to speed up processing of geometric models. Thanks to the idea of progressive interpolation, the compression qualities and noise-filtering effects have been improved significantly. However, the reconstruction precision fails to be improved further because many model details are removed and the noise-filtering performance decreases greatly while the noise intensity increases gradually. To deal with this dilemma, a unified Catmull-Clark subdivision based biorthogonal wavelet construction with shape control parameters is presented to process 3D models with sharp-feature constraints. By customizing its local orthogonalizing coefficients for different vertex valences of quadrilateral patches, the novel scheme can greatly strengthen the capability of the model's shape control that is vital for data compression, noise-filtering, etc. Combined with the local and in-place lifting operations, the proposed wavelet transform can dramatically decrease the memory consumption and computation complexity. Both theoretical analysis and numerical experiments show that, compared with the state-of-the-art lifting-based solutions, the proposed wavelet transform achieves higher compression ratio, more stable noise-filtering effects and better progressive transmission quality, not only decreasing the Bits/vertex of 3D meshes and improving the PSNR of reconstructed models, but also reducing the time costs of coding and decoding.

In imaging of multilayered object, total focusing method (TFM) combined with ray-tracing appeals to numerous researchers because of its superior image quality. But lots of iterative computation to calculate the incident points greatly increases its processing time. Based on line scan-conversion algorithm and Snell's principle, this paper proposes an efficient method to calculate the incident point. By dividing the image area into multiple regions, the time-consuming iterative calculation of point of incidence (POI) is eliminated. After introducing this method into TFM, the time complexity of Region-Division (RD) TFM is reduced by one order of magnitude. The simulation experiments validate that the imaging speed of RD-TFM is improved dozens of times compared to traditional TFM with ray-tracing. Besides, RD-TFM applies to multilayered objects with both planar and non-planar interface.

In non-destructive testing of three-dimensional media, the total focusing method (TFM) draws attentions of numerous researchers owing to its ability to provide image with superior quality. But its application in multilayered media is limited by the computation complexity of calculating point of incidence (POI) and increased quantity of acquired ultrasonic data. This paper presents a novel 3D-TFM called linear scan-conversion (LSC) 3D-TFM for the imaging of 3D multilayered media, which uses the scan-conversion algorithm of line segment to substitute traditional ray-tracing method to calculate POI. The iterative computations in ray-tracing method can be avoided, which reduces the time complexity by one order of magnitude. The simulation experiment indicates that LSC 3D-TFM speeds up the imaging process 28 times while maintaining the same result.

Fast and accurate motion estimation takes an important place in many fields of computer vision and image processing. Using Radon transform to compute projections of the images along specified directions is an effective way to show the relationship between the 2D image object and its projections and estimate motions between the images. All existing projection-based motion estimation methods without the use of iteration have a severe defect that only five of the six affine parameters can be estimated. There are some other methods that can estimate the six parameters, but most of them are usually based on a certain iterative framework, which is computationally intensive and sensitively dependent on the initial values. In this paper, a novel method based on Radon transform is proposed to estimate all the six affine parameters directly. The relationship in the projection domain between a pair of images connected by an affine motion is studied and a linear model is established, by which all the six affine parameters can be directively found. The employment of a hierarchical framework can produce more accurate results. The experimental results reveal that the proposed method has a much better performance than the state-of-the-art methods in this field.

This paper presents a novel two-step examplar based image in painting approach. Firstly, a coarse version of the input image is acquired in a multi-resolution way through the Digital Wavelet Transform (DWT), and then based on this image, the original missing regions of the input image are in painted with the examplar-based method. In this in painted image, most of the structures and textures in the missing regions can be recovered, but some artifacts may exist in some certain areas that contain edges. To overcome this problem, the second pass of in painting is performed only in these areas. Compared with the existing examplar-based methods, a few new techniques, such as the dynamic search window and the patch structure distance test, are proposed to accelerate the best match searching process. It is shown by the experimental results that the proposed algorithm can achieve plausible results very close to those of the state-of-the-art approaches, but in a much faster way with lower computational cost.

High-quality indirect lighting at interactive speed is a difficult challenge. To fast approximate the indirect illumination, the volume-based rendering techniques were used. The Light Propagation Volume (LPV) [Kaplanyan et al. 2010] method departs scenes into coarse lattices and propagates spherical harmonics represented radiance on them. The LPV is able to render complex and dynamic scenes in real-time. But since the radiance transfer is highly approximated among the lattices, it fails to simulate the indirect lighting between the surfaces in the same lattice. On the other hand, the Voxel-based Global Illumination (VGI) [Kaplanyan et al. 2011] voxelizes the scene into fine voxels. It performs the voxel-based ray marching to find the reflected surface in the near-field. By adopting a 1/4x1/4 sampling, the VGI can render in a speed of 18~28 frame per second. But for complex scenes and real global illumination, it requires huge volume data and a large number of rays which degrades the rendering performance to a speed of 2.2s per frame.

Abstract In recent years, some wavelets constructed via local lifting have been proposed to accelerate the compression of geometric models. Unlike the traditional wavelet‐based compression with global optimisation, their compression ratios are still not sufficient, particularly at low‐bit rates. In this paper, we present a novel efficient wavelet‐based compression approach using the matrix‐valued Loop subdivision on triangular meshes. Based on the local lifting scheme, the computation of the wavelet transform is highly efficient and requires low memory usage. With the shape controls introduced by the matrix‐valued subdivision, our approach can greatly improve the fitting quality of surfaces while keeping the efficient data compression. The experimental results showed that our wavelet transform had better performance on balancing the efficiency and compression ratios. Copyright © 2011 John Wiley & Sons, Ltd.

a novel adaptable accurate way for calculating Polar FFT and Log-Polar FFT is developed in this paper, named Multilayer Fractional Fourier Transform (MLFFT). MLFFT is a necessary addition to the pseudo-polar FFT for the following reasons: It has lower interpolation errors in both polar and log-polar Fourier transforms; it reaches better accuracy with the nearly same computing complexity as the pseudo-polar FFT; it provides a mechanism to increase the accuracy by increasing the user-defined computing level. This paper demonstrates both MLFFT itself and its advantages theoretically and experimentally. By emphasizing applications of MLFFT in image registration with rotation and scaling, our experiments suggest two major advantages of MLFFT: 1) scaling up to 5 and arbitrary rotation angles, or scales up to 10 without rotation can be recovered by MLFFT while currently the result recovered by the state-of-the-art algorithms is the maximum scaling of 4; 2) No iteration is needed to obtain large rotation and scaling values of images by MLFFT, hence it is more efficient than the pseudopolar-based FFT methods for image registration.

Abstract In this paper a biorthogonal wavelet approach based on Doo‐Sabin subdivision is presented. In the dual subdivision like Doo‐Sabin scheme, all the old control vertices disappear after one subdivision step, which is a big challenge to the biorthogonal wavelet construction. In our approach, the barycenters of the V‐faces corresponding to the old vertices are selected as the vertices associated with the scaling functions to construct the scaling space. The lifting scheme is used to guarantee the fitting quality of the wavelet transform, and a local orthogonalization is introduced with a discrete inner product operation to improve the computation efficiency. Sharp feature modeling based on extended Doo‐Sabin subdivision rules is also discussed in the framework of our wavelet construction. The presented wavelet construction is proven to be stable and effective by the experimental results.

Reflections, refractions, and caustics are very important for rendering global illumination images. Although many methods can be applied to generate these effects, the rendering performance is not satisfactory for interactive applications. In this paper, complex ray-object intersections are simplified so that the intersections can be computed on a GPU, and an iterative computing scheme based on the depth buffers is used for correcting the approximate results caused by the simplification. As a result, reflections and refractions of environment maps and nearby geometry can be rendered on a GPU interactively without preprocessing. We can even achieve interactive recursive reflections and refractions by using an object-impostor technique. Moreover, caustic effects caused by reflections and refractions can be rendered by placing the eye at the light. Rendered results prove that our method is sufficiently efficient to render plausible images interactively for many interactive applications.

A new efficient biorthogonal wavelet analysis based on the principal square root of subdivision is proposed in the paper by using the lifting scheme. Since the principal square root of subdivision is of the slowest topological refinement among the traditional triangular subdivisions, the multiresolution analysis based on the principal square root of subdivision is more balanced than the existing wavelet analyses on triangular meshes, and accordingly offers more levels of detail for processing polygonal models. In order to optimize the multiresolution analysis process, the new wavelets, no matter whether they are interior or on boundaries, are orthogonalized with the local scaling functions based on a discrete inner product with subdivision masks. Because the wavelet analysis and synthesis algorithms are actually composed of a series of local lifting operations, they can be performed in linear time. The experiments demonstrate the efficiency and stability of the wavelet analysis for both closed and open triangular meshes with principal square root of subdivision connectivity. The principal square root of -subdivision-based biorthogonal wavelets can be used in many applications such as progressive transmission, shape approximation, multiresolution editing and rendering of 3D geometric models.

In this paper, we propose a new wavelet scheme for loop subdivision surfaces. The main idea enabling our wavelet construction is to extend the subdivision rules to be invertible, thus executing each inverse subdivision step in the reverse order makes up the wavelet decomposition rule. As opposed to other existing wavelet schemes for loop surfaces, which require solving a global sparse linear system in the wavelet analysis process, our wavelet scheme provides efficient (linear time and fully in-place) computations for both forward and backward wavelet transforms. This characteristic makes our wavelet scheme extremely suitable for applications in which the speed for wavelet decomposition is critical. We also describe our strategies for optimizing free parameters in the extended subdivision steps, which are important to the performance of the final wavelet transform. Our method has been proven to be effective, as demonstrated by a number of examples.

In computer graphics and computer-aided geometric design, more and more subdivision schemes are being extensively used for free-form surfaces of arbitrary topology. The convergence and continuity analyses of uniform subdivision surfaces have been performed very well, but it is very difficult to prove the convergence and the continuity properties of non-uniform recursive subdivision surfaces (NURSSes, for short) because the subdivision matrix varies at each iteration step. This restricts widespread use of NURSSes, although NURSSes have a lot of advantages over uniform subdivision surfaces. In this paper, the concept of equivalent knot spacing is presented. A new technique for eigenanalysis, convergence and continuity analyses of non-uniform Doo-Sabin surfaces is proposed. Also, an interesting and important fact is found that the subdivision process of nonuniform Doo-Sabin surfaces may diverge sometime.

A new kind of NURBS representation for periodic curves is introduced. The method for periodic curves is extended to derive a compact and symmetric representation of the sphere or the ellipsoid using a piecewise bi-quadratic NURBS surface with only eight distinct control vertices.

Abstract In this paper, a new method is presented for representing conics precisely using NURBS of degree two. The method requires only four control vertices to represent a full circle using the full circumscribing square‐based representation, which is less than the number required with the other methods using the square control vertex configuration.9, 10, 13 The paper also derives a new condition for NURBS to represent conics precisely and shows the sufficient and necessary conditions for representation of a circular arc, and gives the definition and the knot vector formulae of a periodic NURBS curve.