• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Reconstructing piecewise planar scenes with multi-view regularization

    2019-02-27 10:36:48WeijieXiXuejinChen
    Computational Visual Media 2019年4期

    Weijie Xi, Xuejin Chen()

    Abstract Reconstruction of man-made scenes from multi-view images is an important problem in computer vision and computer graphics. Observing that manmade scenes are usually composed of planar surfaces,we encode plane shape prior in reconstructing man-made scenes. Recent approaches for single-view reconstruction employ multi-branch neural networks to simultaneously segment planes and recover 3D plane parameters.However, the scale of available annotated data heavily limits the generalizability and accuracy of these supervised methods. In this paper, we propose multiview regularization to enhance the capability of piecewise planar reconstruction during the training phase, without demanding extra annotated data.Our multi-view regularization enables the consistency among multiple views by making the feature embedding more robust against view change and lighting variations.Thus, the neural network trained by multi-view regularization performs better on a wide range of views and lightings in the test phase. Based on more consistent prediction results, we merge the recovered models from multiple views to reconstruct scenes.Our approach achieves state-of-the-art reconstruction performance compared to previous approaches on the public ScanNet dataset.

    Keywords scene modeling;multi-view;regularization;neural network

    1 Introduction

    Multi-view reconstruction has been extensively studied in computer graphics and computer vision for decades. Studying this problem can be beneficial for many practical applications, such as indoor navigation, augmented reality, and human—robot interaction. However, reconstruction of indoor scenes is non-trivial due to object clutters, occlusions, large variety of object appearances, etc.

    Traditional multi-view reconstruction methods build dense correspondence between pixels and recover the 3D positions of points. Hand-crafted similarity metrics (e.g., normalized cross-correlation and semi-global matching [1, 2]) are employed to compute point-wise correspondences among multiple views. Without integrating contextual information or global geometric prior, these methods have trouble in reconstructing complete and comprise models for scenes with repetitive patterns or many textureless regions.

    With the rapid development of deep learning techniques, many deep neural network-based (DNNbased) methods have been proposed [3—6]. These methods usually compute the pixel-wise plane-sweep cost volume [7] from multiple view images and use the 3D convolutional neural network (CNN) to infer the depth map directly. These methods work well on reconstructing an object where the images are taken under multiple views around the object. However,the computational cost of using 3D CNNs to infer depth map from cost volumes is huge.

    Man-made scenes, especially indoor scenes, are usually composed of plane surfaces [8]. This shape prior provides more global geometric constraint for depth inference and helps on generating cleaner scene models. Recently, a series of deep learning-based methods have been proposed to reconstruct piece-wise planar scenes from image [9—12]. From a single RGB image, these methods simultaneously segment plane instances and estimate 3D plane parameters. The planarity hypothesis has been proven very effective in reconstructing man-made scenes [8, 13].

    Although these deep learning methods have achieved promising success in single-view reconstruction,inadequate consideration of multi-view consistency and the deficiency of densely annotated training data lead to poor generalizability of the inference feature, especially when the viewpoint changes significantly. As a result, over-segmentation and under-segmentation often appear in the predicted plane instance, which gives rise to the wrong reconstruction of the 3D scene. One possible way to solve this problem is to increase the amount of training data. However, it is nontrivial to obtain accurate depth maps and plane annotations.

    In this paper, we propose a multi-view regularization method which enhances existing single-view reconstruction networks for piecewise planar scenes to obtain more robust feature representation. Our multi-view regularization mainly works in the training phase. From multiple images of the same scene, we enforce the network to embed the inference features of the same plane instance under different views as the same as possible while the features of different plane instances are as different as possible. In the test phase, our method is similar to the existing singleview planar reconstruction network. Each single-view image is fed to the network to obtain an instance segmentation map of the plane and corresponding plane parameters to infer the scene depth map. Then the prediction results of multiple images are merged to compose the 3D scene, as shown in Fig. 1. Compared with the existing methods, our method does not bring additional calculation while improving network performance. During the training phase, there is no increase in the annotation amount of the training data.

    Fig. 1 Given multiple images under different viewpoints, we reconstruct the scene by recovering plane segments and the depth map under the planar shape constraint for each image. Based on our multi-view regularization, the recovered planes and depths are more consistent and can be composed to reconstruct the scene more completely.

    In summary, our contributions are the following:

    ·We propose multi-view regularization to enhance the feature extraction capability of existing single-view planar reconstruction networks,especially against to viewpoint changes and lighting variance.

    ·Our method does not require extra annotated training data and does not bring additional computational cost during the test phase.

    ·Our method achieves state-of-the-art planar reconstruction performance on the public dataset ScanNet [14].

    2 Related work

    Multi-view reconstruction. Conventional methods[1,2,15,16]adopt plane-sweep algorithm to construct matching confidence, then optimize the disparity of reference frustum from matching confidence.Although these methods show successful results under ideal Lambertian scenarios [3], they suffer from poor generalizability to common scenes. Incomplete and incorrect reconstruction usually occur due to low texture, object occlusion, or transparency. Recently,deep learning techniques widely success in image recognition and detection, and many deep learningbased methods [3—6] have been proposed for 3D reconstruction. These methods construct a pixel-wise cost volume according to extracted 2D features, then use 3D CNN to directly regress the disparity of reference frustum. While these methods achieve successful performance in object reconstruction benchmark[17,18],they are computationally expensive.

    Planar scene modeling. For 3D plane modeling tasks, traditional methods extract 3D geometric cues and apply global geometric constraints in modeling process. Delage et al. [19] first extract line segments from the input image, and then use Markov Random Field (MRF) model to label superpixels according to the predefined plane classes. But this needs to be established under the assumption of the Manhattan world. Barinova et al. [20] first extract geometric primitives from an input 2D image, and then use the Conditional Random Field (CRF) model to label the extracted geometric primitives. Similarly,assumption that the scene is composed of flat ground and vertical walls should be made. Although these methods can effectively extract and model the plane in the scene, they all need strong prior assumptions which severely restricts the generalizability of the algorithm.

    Learning-based methods.In the literature, a series of learning-based methods have been proposed for plane scene modeling. Early work[21]introduces a learning-based framework to infer the depth map from a single RGB image. With the improvement of deep learning technology, more and more methods based on Convolutional Neural Networks(CNNs)have been proposed. However, most of these techniques simply produce the scene depth map without extracting plane information such as plane instance and plane parameters.

    Recently, several CNN-based methods have been proposed to extract 3D plane structure. Among them,PlaneNet[9]and Yang and Zhou[10]use the semantic segmentation framework to extract plane instances in the scene, and directly infer plane parameters by a CNN. The difference between them is the way of supervision of the plane parameters. PlaneNet [9]uses traditional plane fitting methods to generate annotations for plane parameters, while Yang and Zhou [10] translate the plane parameter prediction problem into plane depth prediction of the scene and use the ground truth depth values to supervise the plane parameters. However, a drawback of these methods is that only a fixed number of planes can be inferred each time because of the widely used semantic segmentation network architecture. This problem was later solved by PlaneRCNN [11] and Yu et al. [12]. Based on the framework of Mask-RCNN, PlaneRCNN [11] first detects the plane in the image and then performs instance segmentation on the detected mask. Yu et al. [12] address this problem by mean-shift clustering of the extracted feature maps to perform instance segmentation for the planes in the scene.

    3 Methodology

    Our multi-view regularization method is proposed to enhance piece-wise planar scene reconstruction from a single RGB image. Combining the recovered 3D planes from multiple images, we could obtain a more complete model for the scene. The network framework adopted by our method is the single-view plane modeling network [12] as the backbone for feature inference. We first quickly introduce the network architecture and then introduce our multiview regularization method. Figure 2 illustrates the framework of our proposed method.

    3.1 Single-view planar modeling network

    For the convenience of presentation,we denote SVPNet as the single-view plane network [12]. Taking a single color image as input,SVPNet infers a depth map from 3D plane recovery,and the blue block in Fig.2 shows the inference pipeline. It first exploits an encoder—decoder framework to extract features from the color image,then through three convolution layers to produce a 2-channel plane embedding feature map, pixel-wise plane parameters,and a binary semantic segmentation mask which indicate plane or not, respectively. The extracted plane embedding feature maps are clustered to obtain plane instance by a modified mean shift clustering algorithm. Concurrently, the scene depth map of plane regions is inferred from extracted plane parameters. Combining the instance-wise plane segmentation mask and inferred depth map, a piecewise 3D planar scene is modeled.

    3.2 Multi-view regularization

    The purpose of our multi-view regularization design is enforcing the consistency of plane feature embedding from different views against view change and lighting variance during the training phase. To enhance singleview planar scene modeling performance, we take multiple images under different views of the same scene as input for training. Among these images, one image is labeled with ground truth plane instance segmentation labelsS, plane parametersand scene depthD. The image with annotation we called reference imageIrand the otherKimages are called source-view images. All of these images are fed intoK+1 share weight sub-networks which include encoder—decoder block and embedding layer which to extract plane embedding featurefor each input color image.

    To establish the pixel-wise correspondence between the reference image andKsource-view images,the embedding feature maps in the source-viewsare projected into the reference view according to their camera parameters and ground truth depth map. Therefore, we obtainK+ 1 projected embedding feature maps from theKsource views and reference view, denote asIn order to formulate the multi-view regularization term, the discriminative loss function is extended from Ref.[22]to multi-view feature maps. The singleview discriminative loss function is defined as

    where the variance termLvarpulls each embedding to the mean embedding of the corresponding plane instance, and the distance termLdistpushes the embedding centers of each plane instances away from each other. The definition of them are

    whereZC=C(C -1) is a normalization constant acting as the average of the push distance between different plane instances,f(x)=max(x,0),Cis the number of the ground truth plane instances,Ncis the number of pixels in the plane clusterc,μcis the mean embedding of the plane clusterc, and{μc}Cis the overall set of mean embedding,xiis theithembedding vector on plane clustercof the embedding featureX,‖ · ‖is the L2 distance term,δvandδdare the two margins forLvarandLdistrespectively.

    The single-view loss is used to constrain the compactness of the embedding space, while our multi-view regularization term tends to enforce the consistency of the embedding vectors of the same plane instance under different views. For each sourceview image, we get the projected embedding feature mapand the corresponding plane instance labelSrwhich is the annotation under reference view image.Combining these projected embedding feature maps with the embedding feature map of the reference view,we denote the asFor each plane instance that can be extracted fromSr, we compute the multi-view regularization term.Specifically, we calculate the mean embeddingμcof all the pixels in each plane instance among all theK+1 embedding feature maps{Fk}K+1. With the shared mean embeddingμcamong allK+1 views,our multi-view regularization term is given by

    Since the ultimate goal is to recover the plane instances and their geometry, we combine the semantic segmentation lossLs, plane parameter lossLnwhich are defined in Ref. [12] with our multi-view regularization termLmvto train our designed model.Therefore, the overall loss function of the network is written as

    All these loss function terms are differentiable, so the entire network can be trained in an end-to-end manner.

    4 Experiments

    We first evaluate our multi-view regularization term on the task of single view reconstruction both quantitatively and qualitatively. Then we show a series of reconstruction results by merging the inferred geometry from multiple views. We implement our method with PyTorch [23]. We use SGD optimizer[24] with a learning rate of 10-4and a weight decay of 10-5, and the batch size is set to 16.

    We construct the dataset for training and test based on the ScanNet [14]. Following previous single-view piecewise planar scene modeling works [9, 12], we first construct a dataset containing 48,214 reference images for training. For each reference image,the ground truth annotation with plane instance segmentation map and plane parameters are obtained by fitting planes on the reconstructed mesh in the scenes of ScanNet. To train the network with the multi-view regularization,we select two images which are 20 frames before and after each reference image in the image sequence with corresponding camera intrinsic and extrinsic parameters. At the test phase of single view reconstruction, the trained network takes only one color image as input. For fairness, we evaluate our model on the test set which consists of 760 images, the same as Refs. [9, 12].

    4.1 Quantitative evaluation

    We first compare our model with the baseline model SVPNet [12]. We evaluate the planar scene modeling performance with the per-plane recall and per-pixel recall metrics which are the percentage of correct prediction of the planes and pixels,respectively. The predicted plane is considered a correct prediction when plane instance segmentation intersection-over-union (IOU) score with the ground truth segmentation map is larger than 0.5 and the mean difference between the inferred depth and ground truth depth map is less than a thresholdσd,which ranges from 0.05 to 0.6 m. The experimental comparison results are shown in Table 1. We can see that our model achieves better planar scene modeling performance than the baseline model,especially when the depth threshold is small. Considering that our method does not bring any additional network parameters and computational cost compared to the baseline model at the test phase, we can conclude that the proposed multi-view regularization improves the capability of the single-view inference network.

    Furthermore, we compare our method with PlaneRCNN [11], PlaneNet [9], and NYU-Toolbox[25]. Figure 3 shows the comparison results the test set. Our method achieves the best performance among these methods, which ulteriorly demonstrates the effectiveness of our method.

    4.2 Qualitative evaluation

    Fig. 3 Comparison between the enhanced SVPNet by our multi-view regularization term and three existing methods for single-view planar scene reconstruction on the ScanNet test set.

    Table 1 Per-pixel and per-plane recalls on the ScanNet test dataset compared with the baseline SVPNet [12]

    Fig. 4 Comparison between the single-view reconstruction results of our method and the baseline SVPNet [12]. (a) Input images. (b) The plane clustering results by SVPNet and (c) our method. (d) The inferred depth map by SVPNet and (e) our method. (f, g) Two novel views of the recovered models using SVPNet and (h, i) our method.

    Figure 4 shows a group of results of plane instance segmentation, depth estimation, and reconstructed models using our method and the baseline model[12] on the ScanNet dataset [14]. Comparing (b)and (c) in Fig. 4, we can observe that our method enhances the ability of the network of detecting and recovering small plane instances and improves the model robustness to avoid over-segmentation of entire plane into multiple pieces. From Figs. 4(d)—4(i), we conclude that our method also enhances the ability of the network to recover correct 3D plane parameters because the three branches of the network share the same feature extractor and they can be mutually improved by joint training.

    4.3 Multiple view consistency evaluation

    The goal we designed the multi-view regularization is to enhance the feature consistency of multiple views, thereby improving the robustness of the network against view change and lighting variance.We validate the multi-view consistency of our method from two perspectives: plane segmentation consistency and 3D planar model correctness.

    Plane segmentation consistency. We randomly select three images from different views in the same scene and test our enhanced single-view reconstruction network on each image independently.Figure 5 displays the plane instance segmentation results obtained by our method and three stateof-the-art approaches [9, 11, 12]. It shows that our method achieves the best result in maintaining the segmentation consistency among different views.Compared with the existing methods, our method can effectively reduce over-segmentation and undersegmentation as highlighted by the orange boxes.Moreover, the plane boundaries generated by our methods are more clear and consistent.

    Geometry correctness. For the task of modeling piecewise planar scenes, the ultimate goal is to get a complete and correct 3D model of the scene. Given multiple images taken under different views in the same scene, we first run our enhanced network to reconstruct the plane instances for each image. Then we transform the reconstructed mesh from each image to the reference view according to their associated camera parameters and merge the meshes to obtain a more complete reconstruction result of the scene.Figure 6 shows the reconstructed results of two scenes using our method and the baseline model [12]. For each scene, we select three frames under adjacent viewpoints for reconstruction. Generally, our method reconstructs more view-consistent results. In the first scene, the red book flys away in the reconstructed model using SVPNet because it is segmented to the wall region by mistake. The desktop area is stretched a lot using SVPNet [12]. In comparison, the shape of the desktop is well recovered and consist among different views. In the second scene, the wall region reconstructed from different views using our method is much more consistent than that reconstructed using SVPNet [12]. These examples illustrate that our multi-view regularization can enhance the robustness of the single-view reconstruction network against view change and lighting variation, and therefore a more accurate 3D scene model can be obtained.

    Fig. 5 Comparison of the plane segmentation results of multiple images under different views using our method and other approaches.

    Fig. 6 The reconstruction results of two scenes from multiple images using our method and SVPNet [12].

    5 Conclusions

    In this paper, we propose a novel method to enhance a single-view reconstruction network by multi-view regularization for modeling piecewise 3D planar scenes. Our method enforces the consistency of the embedding features during the training phase,thereby enhances the robustness of the network against view change and lighting variation. Our method achieves state-of-the-art performance on the task of indoor scene reconstruction on the public ScanNet dataset. We believe that our proposed multiview regularization can be flexibly integrated with many single-view inference networks without bringing extra computational cost.

    Acknowledgements

    This work was supported by the National Key R&D Program of China under Grant 2017YFB1002202,the National Natural Science Foundation of China(NSFC) under Grant 61632006, as well as the Fundamental Research Funds for the Central Universities under Grants WK3490000003 and WK2100100030.

    Open AccessThis article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format,as long as you give appropriate credit to the original author(s)and the source, provide a link to the Creative Commons licence, and indicate if changes were made.

    The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use,you will need to obtain permission directly from the copyright holder.

    To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

    Other papers from this open access journal are available free of charge from http://www.springer.com/journal/41095.To submit a manuscript, please go to https://www.editorialmanager.com/cvmj.

    可以在线观看毛片的网站| 天堂av国产一区二区熟女人妻| 国产乱人偷精品视频| 亚洲av不卡在线观看| 午夜免费激情av| 免费观看人在逋| 综合色av麻豆| 最近视频中文字幕2019在线8| 亚洲欧洲日产国产| 偷拍熟女少妇极品色| 噜噜噜噜噜久久久久久91| АⅤ资源中文在线天堂| 久久久久精品久久久久真实原创| 国产真实伦视频高清在线观看| 日韩欧美精品v在线| 少妇被粗大猛烈的视频| av免费观看日本| 久久久久精品久久久久真实原创| 久久久成人免费电影| 美女国产视频在线观看| 黑人高潮一二区| 亚洲国产最新在线播放| 久久久a久久爽久久v久久| 人人妻人人澡人人爽人人夜夜 | 日韩中字成人| av卡一久久| 国产v大片淫在线免费观看| 成人一区二区视频在线观看| 亚洲性久久影院| 看十八女毛片水多多多| 三级男女做爰猛烈吃奶摸视频| 国产av码专区亚洲av| 国产极品天堂在线| 婷婷色综合大香蕉| 99热网站在线观看| 中国国产av一级| 一区二区三区四区激情视频| 一个人观看的视频www高清免费观看| 欧美激情在线99| 六月丁香七月| 丰满少妇做爰视频| 中文乱码字字幕精品一区二区三区 | 亚洲,欧美,日韩| 高清视频免费观看一区二区 | 赤兔流量卡办理| 欧美高清成人免费视频www| 搞女人的毛片| 欧美区成人在线视频| 欧美精品一区二区大全| 国产三级在线视频| www.色视频.com| 高清日韩中文字幕在线| 色综合站精品国产| 日韩成人av中文字幕在线观看| 99国产精品一区二区蜜桃av| 欧美一区二区亚洲| 亚洲五月天丁香| 国产片特级美女逼逼视频| 青春草国产在线视频| 国产精品蜜桃在线观看| 日本一本二区三区精品| 波多野结衣巨乳人妻| 国产精品一区二区三区四区免费观看| 精品不卡国产一区二区三区| 男女啪啪激烈高潮av片| 亚洲国产精品专区欧美| 国产成人精品婷婷| 欧美成人一区二区免费高清观看| 男女那种视频在线观看| 亚洲国产高清在线一区二区三| 国产午夜福利久久久久久| 亚洲五月天丁香| 最后的刺客免费高清国语| 亚洲精品国产av成人精品| 女人被狂操c到高潮| 又爽又黄a免费视频| 黄色配什么色好看| 身体一侧抽搐| 深夜a级毛片| 精品久久久久久久久亚洲| 美女国产视频在线观看| 亚洲av福利一区| 最近视频中文字幕2019在线8| av视频在线观看入口| 毛片女人毛片| 乱系列少妇在线播放| 久久久久久久国产电影| 国产亚洲av片在线观看秒播厂 | 欧美成人一区二区免费高清观看| 亚洲成人久久爱视频| 国产又色又爽无遮挡免| 国内揄拍国产精品人妻在线| 日韩高清综合在线| 国产激情偷乱视频一区二区| 自拍偷自拍亚洲精品老妇| 国产精品女同一区二区软件| 色尼玛亚洲综合影院| av.在线天堂| 亚洲av成人av| 日本黄色片子视频| 在线观看66精品国产| 丰满人妻一区二区三区视频av| 2021少妇久久久久久久久久久| 亚洲在线观看片| 午夜福利高清视频| 免费观看人在逋| 欧美成人精品欧美一级黄| 看十八女毛片水多多多| 亚洲内射少妇av| 在线免费观看不下载黄p国产| 亚洲人成网站在线观看播放| 国产精品福利在线免费观看| 色5月婷婷丁香| 神马国产精品三级电影在线观看| 国产精品久久久久久av不卡| 看片在线看免费视频| 2022亚洲国产成人精品| 毛片一级片免费看久久久久| 天堂影院成人在线观看| 国产精品一区二区三区四区免费观看| 黄片wwwwww| 青青草视频在线视频观看| 一区二区三区乱码不卡18| 亚洲av成人精品一区久久| 22中文网久久字幕| 欧美性感艳星| 乱码一卡2卡4卡精品| 日韩欧美 国产精品| 久久久久久久午夜电影| 亚洲国产成人一精品久久久| 在线a可以看的网站| 又粗又硬又长又爽又黄的视频| 非洲黑人性xxxx精品又粗又长| 尾随美女入室| 久久精品国产自在天天线| 一区二区三区免费毛片| 国产又色又爽无遮挡免| 国产成人免费观看mmmm| 国产乱来视频区| 国产午夜精品论理片| 男女下面进入的视频免费午夜| 一本久久精品| 欧美+日韩+精品| 搡女人真爽免费视频火全软件| 大香蕉97超碰在线| 91精品一卡2卡3卡4卡| 日本午夜av视频| 日韩欧美精品免费久久| 色5月婷婷丁香| 亚洲欧美成人综合另类久久久 | 亚洲精品自拍成人| 一区二区三区免费毛片| 中国美白少妇内射xxxbb| 成人国产麻豆网| 春色校园在线视频观看| 熟妇人妻久久中文字幕3abv| 亚洲国产欧美在线一区| 国产v大片淫在线免费观看| 插逼视频在线观看| 视频中文字幕在线观看| 联通29元200g的流量卡| www日本黄色视频网| 能在线免费看毛片的网站| 国产淫语在线视频| 亚洲精品成人久久久久久| 午夜老司机福利剧场| 秋霞伦理黄片| 久久99热这里只有精品18| 99在线视频只有这里精品首页| 国产成人午夜福利电影在线观看| 亚洲伊人久久精品综合 | h日本视频在线播放| 最近最新中文字幕免费大全7| 精品一区二区三区视频在线| www.av在线官网国产| 久久婷婷人人爽人人干人人爱| 亚洲av免费高清在线观看| 国产精品人妻久久久久久| av国产免费在线观看| 18禁裸乳无遮挡免费网站照片| 免费观看精品视频网站| 亚洲成人精品中文字幕电影| 2021少妇久久久久久久久久久| 国产又色又爽无遮挡免| 我的女老师完整版在线观看| 干丝袜人妻中文字幕| 三级男女做爰猛烈吃奶摸视频| 高清av免费在线| 日本黄色片子视频| av福利片在线观看| 乱人视频在线观看| 国产又黄又爽又无遮挡在线| 亚洲精品自拍成人| 91久久精品国产一区二区成人| 国产综合懂色| 村上凉子中文字幕在线| 亚洲欧美日韩高清专用| 精品酒店卫生间| 国产男人的电影天堂91| 国产91av在线免费观看| 日本免费一区二区三区高清不卡| 久久99热这里只频精品6学生 | 久久亚洲精品不卡| 99热全是精品| 午夜福利在线观看免费完整高清在| 91久久精品国产一区二区三区| 免费黄网站久久成人精品| 久久精品91蜜桃| 久久久精品欧美日韩精品| 亚洲国产精品国产精品| 亚洲成人中文字幕在线播放| 久久人人爽人人片av| 亚洲欧美日韩东京热| 亚洲精品久久久久久婷婷小说 | 久久6这里有精品| 国产视频内射| 大又大粗又爽又黄少妇毛片口| 国产精品一区二区三区四区免费观看| 永久免费av网站大全| 亚洲国产精品sss在线观看| 亚洲av男天堂| 亚洲欧美日韩卡通动漫| 国产亚洲一区二区精品| 亚洲三级黄色毛片| 人妻夜夜爽99麻豆av| 美女国产视频在线观看| av视频在线观看入口| 亚洲国产成人一精品久久久| 成人欧美大片| 麻豆久久精品国产亚洲av| 九九久久精品国产亚洲av麻豆| 日本爱情动作片www.在线观看| 欧美高清性xxxxhd video| 日韩人妻高清精品专区| 亚洲高清免费不卡视频| 国产探花极品一区二区| 亚洲国产欧美人成| 国产成人a区在线观看| 美女大奶头视频| 老司机影院成人| 国产午夜精品论理片| 亚洲自偷自拍三级| 尤物成人国产欧美一区二区三区| 3wmmmm亚洲av在线观看| 精品无人区乱码1区二区| 免费av不卡在线播放| 中国国产av一级| 久久久久久久亚洲中文字幕| 亚洲国产精品专区欧美| 麻豆av噜噜一区二区三区| 亚洲av日韩在线播放| 女人十人毛片免费观看3o分钟| 可以在线观看毛片的网站| 亚洲欧美成人综合另类久久久 | 亚洲精品日韩av片在线观看| 99热全是精品| 永久免费av网站大全| or卡值多少钱| 国产探花在线观看一区二区| 国产不卡一卡二| 国产人妻一区二区三区在| 人人妻人人澡人人爽人人夜夜 | 日韩成人av中文字幕在线观看| 天堂网av新在线| 男人舔女人下体高潮全视频| 男人狂女人下面高潮的视频| 2021天堂中文幕一二区在线观| 99在线视频只有这里精品首页| 黄色日韩在线| 一二三四中文在线观看免费高清| 欧美人与善性xxx| 国产成人福利小说| 亚洲在线观看片| 男女下面进入的视频免费午夜| 免费在线观看成人毛片| 久久久久久九九精品二区国产| 成人无遮挡网站| 国产又色又爽无遮挡免| 丰满乱子伦码专区| 又黄又爽又刺激的免费视频.| av国产免费在线观看| 51国产日韩欧美| 国产熟女欧美一区二区| 国产精品久久久久久av不卡| 国产毛片a区久久久久| 日韩强制内射视频| 成人午夜精彩视频在线观看| 亚洲av成人av| 精品久久久噜噜| 菩萨蛮人人尽说江南好唐韦庄 | 日韩av在线免费看完整版不卡| 亚洲va在线va天堂va国产| 亚洲欧美日韩东京热| 亚洲最大成人av| 成人综合一区亚洲| 中文精品一卡2卡3卡4更新| 内地一区二区视频在线| 国产不卡一卡二| 美女高潮的动态| 午夜福利网站1000一区二区三区| 老司机影院成人| 欧美日韩综合久久久久久| 日本爱情动作片www.在线观看| 国产又色又爽无遮挡免| 国产视频首页在线观看| 国产激情偷乱视频一区二区| 国产成人福利小说| 国产精品野战在线观看| 国产成人福利小说| av专区在线播放| 亚州av有码| 99久久人妻综合| 五月伊人婷婷丁香| 中文字幕制服av| 久久99蜜桃精品久久| 国产成人aa在线观看| 精品一区二区三区人妻视频| 日本欧美国产在线视频| 天堂中文最新版在线下载 | 99热网站在线观看| 精品久久久久久久久久久久久| 深夜a级毛片| 麻豆精品久久久久久蜜桃| 大话2 男鬼变身卡| 日本爱情动作片www.在线观看| 老女人水多毛片| 国产免费一级a男人的天堂| 舔av片在线| 99久久人妻综合| 色网站视频免费| 免费播放大片免费观看视频在线观看 | av卡一久久| 小蜜桃在线观看免费完整版高清| eeuss影院久久| 成人无遮挡网站| 国产欧美日韩精品一区二区| 久久久久久九九精品二区国产| 国产亚洲最大av| 国产精品久久电影中文字幕| 国产毛片a区久久久久| 国产成人精品婷婷| 午夜爱爱视频在线播放| 中文资源天堂在线| 免费观看a级毛片全部| 久久午夜福利片| 国产欧美日韩精品一区二区| 久久久亚洲精品成人影院| 女人十人毛片免费观看3o分钟| 老师上课跳d突然被开到最大视频| 午夜福利在线观看免费完整高清在| 乱人视频在线观看| 欧美成人午夜免费资源| 两个人的视频大全免费| 简卡轻食公司| 神马国产精品三级电影在线观看| 成年免费大片在线观看| av在线观看视频网站免费| 少妇熟女欧美另类| 国产午夜精品久久久久久一区二区三区| 免费黄色在线免费观看| 波多野结衣高清无吗| 午夜激情欧美在线| 人人妻人人澡欧美一区二区| 午夜爱爱视频在线播放| 午夜日本视频在线| 青春草国产在线视频| av在线观看视频网站免费| 免费在线观看成人毛片| av在线观看视频网站免费| 三级国产精品片| 汤姆久久久久久久影院中文字幕 | 国产精品蜜桃在线观看| 久久精品夜夜夜夜夜久久蜜豆| 欧美另类亚洲清纯唯美| 少妇丰满av| 韩国av在线不卡| 国产高潮美女av| 搡老妇女老女人老熟妇| 婷婷六月久久综合丁香| 男女边吃奶边做爰视频| 亚洲国产精品专区欧美| 建设人人有责人人尽责人人享有的 | 又黄又爽又刺激的免费视频.| 精品国内亚洲2022精品成人| 亚洲伊人久久精品综合 | 边亲边吃奶的免费视频| 嫩草影院入口| 久久久久免费精品人妻一区二区| 我要看日韩黄色一级片| 精品久久久久久久人妻蜜臀av| 亚洲精品乱久久久久久| 十八禁国产超污无遮挡网站| 亚洲av免费在线观看| 女的被弄到高潮叫床怎么办| 久久精品久久久久久噜噜老黄 | 国产精品一及| 亚洲av成人精品一区久久| 美女黄网站色视频| 成人美女网站在线观看视频| 亚洲国产欧美在线一区| 欧美97在线视频| 成人毛片a级毛片在线播放| 波多野结衣高清无吗| 不卡视频在线观看欧美| ponron亚洲| 亚洲18禁久久av| 国产麻豆成人av免费视频| 免费观看精品视频网站| 国产爱豆传媒在线观看| 观看美女的网站| 天堂网av新在线| 精华霜和精华液先用哪个| 亚洲欧美清纯卡通| 97超碰精品成人国产| av播播在线观看一区| 日本免费a在线| 乱系列少妇在线播放| 午夜视频国产福利| 日本黄色视频三级网站网址| 久久欧美精品欧美久久欧美| 国产极品天堂在线| 黑人高潮一二区| 少妇的逼好多水| АⅤ资源中文在线天堂| 久久人人爽人人爽人人片va| 91久久精品国产一区二区成人| 国产精品女同一区二区软件| 成人特级av手机在线观看| 色哟哟·www| 亚洲四区av| 亚洲中文字幕一区二区三区有码在线看| 午夜福利视频1000在线观看| 精品一区二区三区人妻视频| 乱系列少妇在线播放| 午夜日本视频在线| 韩国高清视频一区二区三区| 91av网一区二区| 国产一级毛片在线| 免费播放大片免费观看视频在线观看 | 嫩草影院新地址| 视频中文字幕在线观看| 麻豆乱淫一区二区| 免费av观看视频| 免费无遮挡裸体视频| 日韩欧美三级三区| 一个人看视频在线观看www免费| 欧美人与善性xxx| 最新中文字幕久久久久| 国产老妇伦熟女老妇高清| 成人鲁丝片一二三区免费| 国产久久久一区二区三区| 国产乱人偷精品视频| 天堂av国产一区二区熟女人妻| 国产精品一区二区三区四区免费观看| 欧美又色又爽又黄视频| 哪个播放器可以免费观看大片| 一级av片app| 欧美bdsm另类| videos熟女内射| 国产成人免费观看mmmm| 国产欧美日韩精品一区二区| 激情 狠狠 欧美| 国产伦在线观看视频一区| 免费观看性生交大片5| 白带黄色成豆腐渣| 男人的好看免费观看在线视频| 国产一级毛片七仙女欲春2| 成人av在线播放网站| 国产三级中文精品| 国产精品伦人一区二区| 久久99热这里只有精品18| 最近最新中文字幕大全电影3| АⅤ资源中文在线天堂| 亚洲性久久影院| 岛国在线免费视频观看| 色网站视频免费| 亚洲四区av| 男人舔奶头视频| 干丝袜人妻中文字幕| 男人舔女人下体高潮全视频| 国产又黄又爽又无遮挡在线| a级毛片免费高清观看在线播放| 久久亚洲国产成人精品v| 69人妻影院| 欧美最新免费一区二区三区| 国内精品宾馆在线| 婷婷色麻豆天堂久久 | 日日啪夜夜撸| 91精品国产九色| 国产伦精品一区二区三区视频9| 国产又黄又爽又无遮挡在线| 男女国产视频网站| 日韩,欧美,国产一区二区三区 | 欧美日韩一区二区视频在线观看视频在线 | 我的老师免费观看完整版| 国产黄色视频一区二区在线观看 | 蜜桃久久精品国产亚洲av| 国产私拍福利视频在线观看| 欧美一级a爱片免费观看看| 久久99热这里只有精品18| 亚洲人成网站高清观看| 亚洲精品国产成人久久av| 国产探花极品一区二区| 日韩欧美精品免费久久| 国内少妇人妻偷人精品xxx网站| 韩国高清视频一区二区三区| 亚洲伊人久久精品综合 | 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 一级毛片我不卡| 国产极品精品免费视频能看的| 99热全是精品| 日韩精品有码人妻一区| 亚洲在线观看片| 一级二级三级毛片免费看| 免费无遮挡裸体视频| 99在线视频只有这里精品首页| 国产精品福利在线免费观看| 亚洲图色成人| 床上黄色一级片| 国产成人一区二区在线| videossex国产| 91久久精品电影网| 女人被狂操c到高潮| 七月丁香在线播放| 亚洲欧美成人精品一区二区| 国产精品久久久久久av不卡| 午夜福利高清视频| 晚上一个人看的免费电影| 国产精品,欧美在线| 午夜视频国产福利| 男女视频在线观看网站免费| 亚洲精品aⅴ在线观看| 久久精品久久久久久久性| 免费av毛片视频| 99久久精品一区二区三区| 搡老妇女老女人老熟妇| 国产成人精品久久久久久| 欧美丝袜亚洲另类| 精华霜和精华液先用哪个| 又黄又爽又刺激的免费视频.| 成人漫画全彩无遮挡| 国产精品1区2区在线观看.| 国产美女午夜福利| 少妇被粗大猛烈的视频| 黄色欧美视频在线观看| 嫩草影院新地址| 国产成人精品久久久久久| 成人鲁丝片一二三区免费| 美女黄网站色视频| 国产伦在线观看视频一区| 久久久久九九精品影院| 久久久午夜欧美精品| 亚洲国产精品国产精品| 国国产精品蜜臀av免费| 黄色一级大片看看| 99在线视频只有这里精品首页| 在线观看66精品国产| 国产成人精品婷婷| 韩国av在线不卡| 波野结衣二区三区在线| 九九久久精品国产亚洲av麻豆| 丝袜美腿在线中文| 成年av动漫网址| 中文字幕av成人在线电影| 又粗又硬又长又爽又黄的视频| 老司机影院成人| 国产精品乱码一区二三区的特点| 精品一区二区免费观看| 欧美xxxx性猛交bbbb| www.av在线官网国产| 亚洲av免费在线观看| 老司机福利观看| 国产女主播在线喷水免费视频网站 | 色网站视频免费| 日本爱情动作片www.在线观看| 欧美97在线视频| 九九久久精品国产亚洲av麻豆| 啦啦啦啦在线视频资源| 成人亚洲欧美一区二区av| 日韩亚洲欧美综合| 91av网一区二区| 成人一区二区视频在线观看| 卡戴珊不雅视频在线播放| 国产三级中文精品| 国产大屁股一区二区在线视频| 青青草视频在线视频观看| 欧美日韩一区二区视频在线观看视频在线 | 午夜a级毛片| 日韩一区二区视频免费看| 一二三四中文在线观看免费高清| 91久久精品电影网| 国产在线一区二区三区精 | 一级毛片aaaaaa免费看小| 毛片一级片免费看久久久久| 日韩一本色道免费dvd| 别揉我奶头 嗯啊视频| 国产精品爽爽va在线观看网站| av天堂中文字幕网| 国产乱来视频区| 亚洲国产欧美在线一区| 女人久久www免费人成看片 | 国产不卡一卡二| 国产精品野战在线观看| 成人无遮挡网站| 2021天堂中文幕一二区在线观| 国产色爽女视频免费观看| 两个人视频免费观看高清| 久久精品人妻少妇| 级片在线观看| 免费看美女性在线毛片视频| 国产一级毛片七仙女欲春2| 精品99又大又爽又粗少妇毛片| 国产精品日韩av在线免费观看| 精品久久久久久久久亚洲| 亚洲丝袜综合中文字幕|