This paper presents HITNet, a novel neural network architecture for real-time stereo matching. Contrary to many recent neural network approaches that operate on a full cost volume and rely on 3D convolutions, our approach does not explicitly build a volume and instead relies on a fast multi-resolution initialization step, differentiable 2D geometric propagation and warping mechanisms to infer disparity hypotheses. To achieve a high level of accuracy, our network not only geometrically reasons about disparities but also infers slanted plane hypotheses allowing to more accurately perform geometric warping and upsampling operations. Our architecture is inherently multi-resolution allowing the propagation of information across different levels. Multiple experiments prove the effectiveness of the proposed approach at a fraction of the computation required by state-of-the-art methods. At the time of writing, HITNet ranks 1st-3rd on all the metrics published on the ETH3D website for two view stereo, ranks 1st on most of the metrics amongst all the end-to-end learning approaches on Middlebury-v3, ranks 1st on the popular KITTI 2012 and 2015 benchmarks among the published methods faster than 100 ms.

/pdf/hitnet-hierarchical-iterative-tile-refinement-network-for-4kge9xy0sp.pdf

HITNet: Hierarchical Iterative Tile Refinement Network for Real-time Stereo Matching

This paper presents HITNet, a novel neural network architecture for real-time stereo matching. Contrary to many recent neural network approaches that operate on a full cost volume and rely on 3D convolutions, our approach does not explicitly build a volume and instead relies on a fast multi-resolution initialization step, differentiable 2D geometric propagation and warping mechanisms to infer disparity hypotheses. To achieve a high level of accuracy, our network not only geometrically reasons about disparities but also infers slanted plane hypotheses allowing to more accurately perform geometric warping and upsampling operations. Our architecture is inherently multi-resolution allowing the propagation of information across different levels. Multiple experiments prove the effectiveness of the proposed approach at a fraction of the computation required by state-of-the-art methods. At the time of writing, HITNet ranks 1st-3rd on all the metrics published on the ETH3D website for two view stereo, ranks 1st on most of the metrics among all the end-to-end learning approaches on Middlebury-v3, ranks 1st on the popular KITTI 2012 and 2015 benchmarks among the published methods faster than 100ms.

The similar analysis of time sequence images to achieve image matching is a foundation of tasks in dynamic environments, such as multi-object tracking and dynamic gesture recognition. Therefore, we propose a matching method of time sequence images based on the Siamese network. Inspired by comparative learning, two different comparative parts are designed and embedded in the network. The first part makes a comparison between the input image pairs to generate the correlation matrix. The second part compares the correlation matrix, which is the output of the first comparison part, with a template, in order to calculate the similarity. The improved loss function is used to constrain the image matching and similarity calculation. After experimental verification, we found that it not only performs better, but also has some ability to estimate the camera pose.

A Time Sequence Images Matching Method Based on the Siamese Network.

We present a novel algorithm called crosspatch-based rolling label expansion for accurate stereo matching. This optimization-based approach can effectively estimate the 3D label of each pixel from huge and infinite label space and then generate a continuous disparity map. The algorithm has two obvious characteristics when compared with the traditional label expansion algorithms. The first feature is the cross-based multilayer structure, where each layer contains a series of cross patches with adaptive shapes, reflecting the edge structure of objects on the image. Besides, such cross patches are non-overlapping and independent, satisfying the submodular property for employing graph cuts. The second feature is the rolling optimization, that firstly generates new label proposal by expanding candidate labels within cross patches, then globally updates labels for the whole image using a proposed rolling move. The experimental results show the high matching accuracy of our method, both in pixel level and subpixel level. According to the latest ranking list of Middlebury 3.0 benchmark, our method is one of the best stereo matching algorithms.

/pdf/crosspatch-based-rolling-label-expansion-for-dense-stereo-2xct89ffj6.pdf

CrossPatch-Based Rolling Label Expansion for Dense Stereo Matching

This paper presents the overview and rationale behind the Decoder-Side Depth Estimation (DSDE) mode of the MPEG Immersive Video (MIV) standard, using the Geometry Absent profile, for efficient compression of immersive multiview video. A MIV bitstream generated by an encoder operating in the DSDE mode does not include depth maps. It only contains the information required to reconstruct them in the client or in the cloud: decoded views and metadata. The paper explains the technical details and techniques supported by this novel MIV DSDE mode. The description additionally includes the specification on Geometry Assistance Supplemental Enhancement Information which helps to reduce the complexity of depth estimation, when performed in the cloud or at the decoder side. The depth estimation in MIV is a non-normative part of the decoding process, therefore, any method can be used to compute the depth maps. This paper lists a set of requirements for depth estimation, induced by the specific characteristics of the DSDE. The depth estimation reference software, continuously and collaboratively developed with MIV to meet these requirements, is presented in this paper. Several original experimental results are presented. The efficiency of the DSDE is compared to two MIV profiles. The combined non-transmission of depth maps and efficient coding of textures enabled by the DSDE leads to efficient compression and rendering quality improvement compared to the usual encoder-side depth estimation. Moreover, results of the first evaluation of state-of-the-art multiview depth estimators in the DSDE context, including machine learning techniques, are presented.

Overview and Efficiency of Decoder-Side Depth Estimation in MPEG Immersive Video

Depth image-based rendering (DIBR) plays an important role in 3D video and free viewpoint video synthesis. However, artifacts might occur in the synthesized view due to viewpoint changes and stereo depth estimation errors. Holes are usually out-of-field regions and disocclusions, and filling them appropriately becomes a challenge. In this paper, a virtual view synthesis approach based on asymmetric bidirectional DIBR is proposed. A depth image preprocessing method is applied to detect and correct unreliable depth values around the foreground edges. For the primary view, all pixels are warped to the virtual view by the modified DIBR method. For the auxiliary view, only the selected regions are warped, which contain the contents that are not visible in the primary view. This approach reduces the computational cost and prevents irrelevant foreground pixels from being warped to the holes. During the merging process, a color correction approach is introduced to make the result appear more natural. In addition, a depth-guided inpainting method is proposed to handle the remaining holes in the merged image. Experimental results show that, compared with bidirectional DIBR, the proposed rendering method can reduce about 37% rendering time and achieve 97% hole reduction. In terms of visual quality and objective evaluation, our approach performs better than the previous methods.

Virtual View Synthesis Based on Asymmetric Bidirectional DIBR for 3D Video and Free Viewpoint Video

The depth image based rendering (DIBR) is a popular technology for 3D video and free viewpoint video (FVV) synthesis, by which numerous virtual views can be generated from a single reference view and its depth image. However, some artifacts are produced in the DIBR process and reduce the visual quality of virtual view. Due to the diversity of artifacts, effectively handling them becomes a challenging task. In this paper, an artifact handling method based on depth image is proposed. The reference image and its depth image are extended to fill the holes that belong to the out-of-field regions. A depth image preprocessing method is applied to project the ghosts to their correct place. The 3D warping process is optimized by an adaptive one-to-four method to deal with the cracks and pixel overlapping. For disocclusions, we calculate depth and background terms of the filling priority based on depth information. The search for the best matching patch is performed simultaneously in the reference image and the virtual image. Moreover, adaptive patch size is used in all hole-filling processes. Experimental results demonstrate the effectiveness of the proposed method, which has better performance compared with previous methods in subjective and objective evaluation.

Artifact Handling Based on Depth Image for View Synthesis

View synthesis is an effective method to generate the contents of multiple views based on a limited number of reference views, which can be used in 2D to 3D conversion, free viewpoint video and multiview video rendering. Depth-image-based rendering (DIBR) is a practical technique to generate virtual view by using a 2D reference view and its depth image. However, a critical problem in DIBR process is that disocclusions might be produced in the synthesized image because the background occluded by the foreground objects in the reference view may be exposed in the virtual view. In this paper, a local foreground removal method is proposed for disocclusion filling. Morphology-based depth image preprocessing is performed before DIBR, aiming to correct the depth value of the ghosts and remove ghost artifacts. In the synthesized virtual image, pixels on the disocclusion edge are identified and classified. Then they are positioned in the reference image by inverse 3D warping. Local foreground regions that occlude the corresponding background are removed from both the reference image and its depth image based on the disocclusion edge pixels. Removed region is filled with surrounding background contents, and depth information is used in this process to prevent foreground penetration. The predicted background contents are warped to the disocclusion region, thereby achieving the hole filling. Experimental results show that the proposed method performs better than the other methods in disocclusion filling, and improves the subjective and objective quality of the synthesized view. In the evaluation results of PSNR, SSIM, FSIMc and VSI, our method improves by 0.32-2.43dB, 0.0036-0.0155, 0.0041-0.0198 and 0.0012-0.0057 respectively compared with competitive methods.

/pdf/local-foreground-removal-disocclusion-filling-method-for-1m4rw7nhwi.pdf

Local Foreground Removal Disocclusion Filling Method for View Synthesis

New virtual view is synthesized through depth image based rendering(DIBR) using a single color image and its associated depth map in 3D view generation. Holes are unavoidably generated in the 2D to 3D conversion process. We propose a hole-filling method based on depth map to address the problem. Firstly, we improve the process of DIBR by proposing a one-to-four (OTF) algorithm. The “z-buffer” algorithm is used to solve overlap problem. Then, based on the classical patch-based algorithm of Criminisi et al., we propose a hole-filling algorithm using the information of depth map to handle the image after DIBR. In order to improve the accuracy of the virtual image, inpainting starts from the background side. In the calculation of the priority, in addition to the confidence term and the data term, we add the depth term. In the search for the most similar patch in the source region, we define the depth similarity to improve the accuracy of searching. Experimental results show that the proposed method can effectively improve the quality of the 3D virtual view subjectively and objectively.

Huaiyuan Xu

Papers

CrossPatch-Based Rolling Label Expansion for Dense Stereo Matching

Virtual View Synthesis Based on Asymmetric Bidirectional DIBR for 3D Video and Free Viewpoint Video

Artifact Handling Based on Depth Image for View Synthesis

Local Foreground Removal Disocclusion Filling Method for View Synthesis

An efficient hole-filling method based on depth map in 3D view generation