[CVPR2020]论文翻译SwapText: Image Based Texts Transfer in Scenes

Posted

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了[CVPR2020]论文翻译SwapText: Image Based Texts Transfer in Scenes相关的知识,希望对你有一定的参考价值。

参考技术A

由于不同因素之间的复杂作用,在保留原始字体,颜色,大小和背景纹理的同时在场景图像中交换文本是一项具有挑战性的任务。在这项工作中,我们提出了一个三阶段框架SwapText,用于跨场景图像传输文本。 首先,提出了一种新颖的文本交换网络来仅替换前景图像中的文本标签。 其次,背景完成网络来学习以重建背景图像。 最后,通过融合网络将生成的前景图像和背景图像用于生成文字图像。 使用提出的框架,即使出现严重的几何失真,我们也可以巧妙的处理输入图像的文本。 定性和定量结果显示在几个场景文本数据集上,包括规则和不规则文本数据集。 我们进行了广泛的实验以证明我们的方法的有效性,例如基于图像的文本翻译,文本图像合成等。

想象一下,能够在场景图像中交换文本,同时在几秒钟内保持原始字体,颜色,大小和背景纹理,而无需花费数小时进行图像编辑。 在这项工作中,我们旨在通过自动替换场景图像中文本的算法来实现此目标。文本交换的核心挑战在于生成视觉逼真的文本并与原始文本保持一致的样式。

文本交换或文本替换在许多情况下都涉及到,包括文本检测,文本识别,海报中的文本转换和其他创造性应用。 对于文本检测和识别任务,文本交换是一种非常有用的数据增强方法。 见证了深度神经网络(DNN)在各种计算机视觉任务中的巨大成功,获得大量带注释的训练图像已成为训练DNN模型的瓶颈。最简单,使用最广泛的方法是通过几何变换来增加训练图像,例如平移,旋转和翻转等。近来,已经提出了基于图像合成的方法[11、7、39]来训练文本检测和识别模型。这些方法通过结合不同的渲染技术对光和能量的物理行为进行建模来从无文本图像中创建新图像。但是, 合成图像无法与场景中的图像完全融合,这在将合成图像应用于DNN模型训练时至关重要。

近年来,许多图像生成模型,例如生成对抗网络(GAN)[6],可变自动编码器(VAE)[17]和自回归模型[25],为现实的图像生成任务提供了强大的工具。在[9,38,33]中,GAN用于图像补全,可为缺失区域生成视觉上逼真的和语义上合理的像素。 [21,8,28,22]已经利用这些网络生成具有不同姿势或服装的新颖人物图像。

我们的贡献总结如下:

文本图像合成
图像合成已在计算机图形学研究中得到了广泛的研究[4]。文本图像合成被研究为一种数据增强方法,用于训练准确而健壮的DNN模型。例如,Jaderberg等[11]使用单词生成器来生成用于文本识别任务的合成单词图像。Gupta等 [7]开发了一个健壮的引擎来生成用于文本检测和识别任务的合成文本图像。 文本图像合成的目标是将文本插入背景图像中语义上敏感的区域。许多因素都影响合成文本图像的真实相似度,例如文本大小,文本视角,环境光照等。 在[39]中,Zhanet等人通过结合语义连贯,视觉注意力和自适应文本外观这三种设计来实现文本文本图像合成。尽管文本图像合成在视觉上是逼真的,但合成图像与真实图像之间仍存在许多差异。例如, 与真实图像相比,合成图像中文本字体和背景图像非常有限。

在最近,基于GAN的图像合成技术得到了进一步的探索。在[41]中,Zhan等人提出了一种将几何合成器和外观合成器组合在一起的空间融合GAN,以在几何和外观空间中实现合成现实。Yang等人[36]使用双向形状匹配框架通过可调整的参数来控制字形的关键风格。 GA-DAN [40]提出了一项有趣的工作,能够同时在几何空间和外观空间中对跨域移位进行建模。[2]中提出了MC-GAN来实现从A到Z的字母集的字体样式转换。 Wu等人 [34]提出了一个端到端的可训练样式保留网络来编辑自然图像中的文本。

图像生成
随着生成模型(例如GAN [6],VAE [17]和自动回归模型[25])的巨大成功,逼真而清晰的图像生成最近吸引了越来越多的关注。传统的生成模型使用GAN [6]或VAE [17]来将噪声z生成的分布映射到实际数据的分布。例如,GANs [6]用于生成真实面孔[37、3、15]和鸟类[29]。

为了控制所生成的结果,Mirzaet等人[23]提出了有条件的GAN。它们会生成在类别标签上进行分类的MNIST数字。在[12]中,karacanet等。根据语义布局和场景属性(例如日夜,晴天雾天)生成逼真的室外场景图像。 Lassneretal [19]基于细粒度的身体和衣服片段生成了穿着者的全身图像。完整模型可以以姿势,形状或颜色为条件。Ma[21,22]基于图像和姿势生成人图像。在[18]中提出了快速人脸交换,以将输入身份转换为目标身份,同时保留姿势,面部表情和光照。

图像完成
最近,基于GAN的方法已经成为图像完成的一种有希望的范例。 Iizuka等 [9]提议使用全局和局部判别器作为对抗性损失,在其中全局和本地一致性都得到了加强。Yu等人 [38]使用上下文注意力层来显式地参与远距离空间位置上的相关特征补丁。 Wang等 [33]使用多列网络以并行方式生成不同的图像分量,并采用隐式的多样化MRF正则化来增强局部细节。

给定场景文本图像Is,我们的目标是在保持原始样式的基础上基于内容图像Ic替换文本。 如图2所示,我们的框架由文本交换网络,背景完成网络和融合网络组成。文本交换网络首先从Is中提取样式特征从Ic中提取内容特征,然后通过自注意网络合并这两个特征。 为了更好地表示内容,我们使用内容形状转换网络(CSTN)根据样式图像Is的几何属性来转换内容图像Ic。背景完成网络用于重建样式图像Is的原始背景图像Ib。 最后,文本交换网络和背景完成网络的输出被融合网络融合以生成最终的文本图像。

现实情况下的文本实例具有多种形状,例如,呈水平,定向或弯曲形式。 文本交换网络的主要目的是在保留原始样式(尤其是文本形状)的同时替换样式图像Is的内容。 为了提高不规则文本图像生成的性能,我们提出了一个内容形状转换网络(CSTN)将内容图像映射到样式图像的相同几何形状中,然后通过3个下采样卷积层和几个残差块对样式图像和转换后的内容图像进行编码。 为了充分融合样式和内容特征,我们将它们馈入了一个自注意网络。 对于解码,使用3个上采样反卷积层来生成前景图像If。

文本形状的定义对于内容形状的转换至关重要。 受文本检测[20]和文本识别[35]领域中的文本形状定义的启发,可以使用2 K个基准点P = p1,p2,...,p2K定义文本的几何尺寸属性,如图3所示。

在对内容和样式图像进行编码之后,我们将两个特征图都馈送到自注意网络,该网络会自动学习内容特征图Fc和样式特征图Fs之间的对应关系。 输出特征图是Fcs,图5(a)给出了自注意力的网络结构。

内容特征Fc和样式特征Fs首先沿其深度轴连接。 然后,我们遵循[42]中类似的自注意力机制来生成输出特征图Fcs。

除了这种单级样式化之外,我们还开发了多级样式化管道,如图5(b)所示。 我们将自注意力网络依次应用于多个特征图层,以生成更逼真的图像。

文本交换网络主要侧重于前景图像生成,而背景图像在最终图像生成中也起着重要作用。为了生成更逼真的文字图像,我们使用背景完成网络来重建背景图像,其结构如表1所示。大多数现有的图像完成方法都是通过借用或复制周围区域的纹理来填充图像的像素。一般的结构遵循编码器-解码器结构,我们在编码器之后使用膨胀卷积层来计算具有较大输入区域的输出像素。通过使用较低分辨率的膨胀卷积,模型可以有效地“看到”输入图像的较大区域。

在此阶段,将文本交换网络和背景完成网络的输出融合以生成完整的文本图像。 如图2所示,融合网络遵循编码器-解码器结构。 类似于[34],我们在融合解码器的上采样阶段将背景完成网络的解码特征图连接到具有相同分辨率的相应特征图。 我们使用Gfuse和Dfuse分别表示生成器和判别器网络。 融合网络的损失函数可计算如下:

为了制作更逼真的图像,我们还遵循样式迁移网络[5,26]的类似思想,将VGG-loss引入融合模块。 VGG损失分为两部分,即知觉损失和风格损失,如下所示:

我们遵循[34]中的类似思想来生成具有相同样式的成对合成图像。我们使用超过1500个字体和10000个背景图像来生成总共100万个狮子训练图像和10000个测试图像。输入图像的大小调整为64×256,批处理大小为32。从权重为零的正态分布初始化所有权重,标准差为0.01。使用β1= 0.9和β2= 0.999的Adam优化器[16]来优化整个框架。在训练阶段将学习率设置为0.0001。我们在Ten-sorFlow框架[1]下实现我们的模型。我们的方法中的大多数模块都是GPU加速的。

我们在几个公共基准数据集上评估了我们提出的方法。

我们采用图像生成中常用的指标来评估我们的方法,其中包括:

在本节中,我们将通过经验研究不同的模型设置如何影响我们提出的框架的性能。我们的研究主要集中在以下方面:内容形状转换网络,自注意力网络和背景完成网络中的膨胀卷积。图6给出了一些定性结果。

自注意力网络
使用自注意力网络来充分结合内容特征和风格特征。根据表2,使用单层自注意力网络,平均l2误差减少约0.003,平均PSNR增加约0.3,平均SSIM增加约0.012。为了使用样式和内容特征的更多全局统计信息,我们采用了一个多层的自注意力网络来融合全局和局部模式。借助多级自我关注网络,所有的度量方法都得到了改进。

膨胀卷积
膨胀卷积层可以扩大像素区域以重建背景图像,因此更容易生成更高质量的图像。 根据表2,具有膨胀卷积层的背景完成网络在所有指标上均具有更好的性能。

为了评估我们提出的方法,我们将其与两种文本交换方法进行了比较:[10]中提出的pix2pix和Wuet等人[34]提出的SRNet。 我们使用生成的数据集来训练和测试这两个模型。根据论文,两种方法都保持相同的配置。

定量结果
在表2中,我们给出了本方法和其他两种竞争方法的定量结果。显然,我们提出的方法在不同语言的所有指标上都有显著改进,平均l2误差减少了0.009以上,平均PSNR增加了0.9以上,平均SSIM增加了0.04以上。第二个最好的方法。

基于图像的翻译是任意文本样式传输的最重要应用之一。在本节中,我们提供一些基于图像的翻译示例,如图7所示。我们在英语和中文之间进行翻译。从结果可以看出,无论目标语言是中文还是英文,都可以很好地保持颜色,几何变形和背景纹理,并且字符的结构与输入文本相同。

在图9中,我们还展示了在场景文本数据集上评估的模型的一些示例结果。根据图9, 我们的模型可以替换输入图像中的文本,同时保留原始字体,颜色,大小和背景纹理。

我们的方法有以下局限性。由于训练数据量有限,因此无法充分利用几何属性空间和字体空间。当样式图像中的文本出现波动时,我们提出的方法将失败,请参见图8(顶部)。图8(底部)显示了使用WordArt中的样式图像的失败案例。

在这项研究中,我们提出了一种健壮的场景文本交换框架SwapText,以解决用预期的文本替换场景文本图像中的文本的新任务。我们采用分而治之的策略,将问题分解为三个子网络,即文本交换网络,背景完成网络和融合网络。在文本交换网络中,内容图像和样式图像的特征被同时提取,然后通过自注意网络进行组合。为了更好地学习内容图像的表示,我们使用内容形状转换网络(CSTN)根据样式图像的几何属性对内容图像进行转换。然后,使用背景完成网络来生成内容图像的背景图像样式图片。最后,将文本交换网络和背景完成网络的输出馈送到融合网络中,以生成更真实和语义一致的图像。在几个公共场景文本数据集上的定性和定量结果证明了我们方法的优越性。在未来的工作中,我们将探索基于字体和颜色生成更多可控制的文本图像。

CVPR2020 论文和代码合集

CVPR2020-Code

CVPR 2020 论文开源项目合集,同时欢迎各位大佬提交issue,分享CVPR 2020开源项目

【推荐阅读】

  • CVPR 2020 virtual

  • ECCV 2020 论文开源项目合集来了:https://github.com/amusi/ECCV2020-Code

  • 关于往年CV顶会论文(如ECCV 2020、CVPR 2019、ICCV 2019)以及其他优质CV论文和大盘点,详见: https://github.com/amusi/daily-paper-computer-vision

【CVPR 2020 论文开源目录】

CNN

Exploring Self-attention for Image Recognition

  • 论文:https://hszhao.github.io/papers/cvpr20_san.pdf

  • 代码:https://github.com/hszhao/SAN

Improving Convolutional Networks with Self-Calibrated Convolutions

  • 主页:https://mmcheng.net/scconv/

  • 论文:http://mftp.mmcheng.net/Papers/20cvprSCNet.pdf

  • 代码:https://github.com/backseason/SCNet

Rethinking Depthwise Separable Convolutions: How Intra-Kernel Correlations Lead to Improved MobileNets

  • 论文:https://arxiv.org/abs/2003.13549
  • 代码:https://github.com/zeiss-microscopy/BSConv

图像分类

Interpretable and Accurate Fine-grained Recognition via Region Grouping

  • 论文:https://arxiv.org/abs/2005.10411

  • 代码:https://github.com/zxhuang1698/interpretability-by-parts

Compositional Convolutional Neural Networks: A Deep Architecture with Innate Robustness to Partial Occlusion

  • 论文:https://arxiv.org/abs/2003.04490

  • 代码:https://github.com/AdamKortylewski/CompositionalNets

Spatially Attentive Output Layer for Image Classification

  • 论文:https://arxiv.org/abs/2004.07570
  • 代码(好像被原作者删除了):https://github.com/ildoonet/spatially-attentive-output-layer

视频分类

SmallBigNet: Integrating Core and Contextual Views for Video Classification

  • 论文:https://arxiv.org/abs/2006.14582
  • 代码:https://github.com/xhl-video/SmallBigNet

目标检测

Overcoming Classifier Imbalance for Long-tail Object Detection with Balanced Group Softmax

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/Li_Overcoming_Classifier_Imbalance_for_Long-Tail_Object_Detection_With_Balanced_Group_CVPR_2020_paper.pdf
  • 代码:https://github.com/FishYuLi/BalancedGroupSoftmax

AugFPN: Improving Multi-scale Feature Learning for Object Detection

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/Guo_AugFPN_Improving_Multi-Scale_Feature_Learning_for_Object_Detection_CVPR_2020_paper.pdf
  • 代码:https://github.com/Gus-Guo/AugFPN

Noise-Aware Fully Webly Supervised Object Detection

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/html/Shen_Noise-Aware_Fully_Webly_Supervised_Object_Detection_CVPR_2020_paper.html
  • 代码:https://github.com/shenyunhang/NA-fWebSOD/

Learning a Unified Sample Weighting Network for Object Detection

  • 论文:https://arxiv.org/abs/2006.06568
  • 代码:https://github.com/caiqi/sample-weighting-network

D2Det: Towards High Quality Object Detection and Instance Segmentation

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/Cao_D2Det_Towards_High_Quality_Object_Detection_and_Instance_Segmentation_CVPR_2020_paper.pdf

  • 代码:https://github.com/JialeCao001/D2Det

Dynamic Refinement Network for Oriented and Densely Packed Object Detection

  • 论文下载链接:https://arxiv.org/abs/2005.09973

  • 代码和数据集:https://github.com/Anymake/DRN_CVPR2020

Scale-Equalizing Pyramid Convolution for Object Detection

论文:https://arxiv.org/abs/2005.03101

代码:https://github.com/jshilong/SEPC

Revisiting the Sibling Head in Object Detector

  • 论文:https://arxiv.org/abs/2003.07540

  • 代码:https://github.com/Sense-X/TSD

Scale-equalizing Pyramid Convolution for Object Detection

  • 论文:暂无
  • 代码:https://github.com/jshilong/SEPC

Detection in Crowded Scenes: One Proposal, Multiple Predictions

  • 论文:https://arxiv.org/abs/2003.09163
  • 代码:https://github.com/megvii-model/CrowdDetection

Instance-aware, Context-focused, and Memory-efficient Weakly Supervised Object Detection

  • 论文:https://arxiv.org/abs/2004.04725
  • 代码:https://github.com/NVlabs/wetectron

Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection

  • 论文:https://arxiv.org/abs/1912.02424
  • 代码:https://github.com/sfzhang15/ATSS

BiDet: An Efficient Binarized Object Detector

  • 论文:https://arxiv.org/abs/2003.03961
  • 代码:https://github.com/ZiweiWangTHU/BiDet

Harmonizing Transferability and Discriminability for Adapting Object Detectors

  • 论文:https://arxiv.org/abs/2003.06297
  • 代码:https://github.com/chaoqichen/HTCN

CentripetalNet: Pursuing High-quality Keypoint Pairs for Object Detection

  • 论文:https://arxiv.org/abs/2003.09119
  • 代码:https://github.com/KiveeDong/CentripetalNet

Hit-Detector: Hierarchical Trinity Architecture Search for Object Detection

  • 论文:https://arxiv.org/abs/2003.11818
  • 代码:https://github.com/ggjy/HitDet.pytorch

EfficientDet: Scalable and Efficient Object Detection

  • 论文:https://arxiv.org/abs/1911.09070
  • 代码:https://github.com/google/automl/tree/master/efficientdet

3D目标检测

SESS: Self-Ensembling Semi-Supervised 3D Object Detection

  • 论文: https://arxiv.org/abs/1912.11803

  • 代码:https://github.com/Na-Z/sess

Associate-3Ddet: Perceptual-to-Conceptual Association for 3D Point Cloud Object Detection

  • 论文: https://arxiv.org/abs/2006.04356

  • 代码:https://github.com/dleam/Associate-3Ddet

What You See is What You Get: Exploiting Visibility for 3D Object Detection

  • 主页:https://www.cs.cmu.edu/~peiyunh/wysiwyg/

  • 论文:https://arxiv.org/abs/1912.04986

  • 代码:https://github.com/peiyunh/wysiwyg

Learning Depth-Guided Convolutions for Monocular 3D Object Detection

  • 论文:https://arxiv.org/abs/1912.04799
  • 代码:https://github.com/dingmyu/D4LCN

Structure Aware Single-stage 3D Object Detection from Point Cloud

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/html/He_Structure_Aware_Single-Stage_3D_Object_Detection_From_Point_Cloud_CVPR_2020_paper.html

  • 代码:https://github.com/skyhehe123/SA-SSD

IDA-3D: Instance-Depth-Aware 3D Object Detection from Stereo Vision for Autonomous Driving

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/Peng_IDA-3D_Instance-Depth-Aware_3D_Object_Detection_From_Stereo_Vision_for_Autonomous_CVPR_2020_paper.pdf

  • 代码:https://github.com/swords123/IDA-3D

Train in Germany, Test in The USA: Making 3D Object Detectors Generalize

  • 论文:https://arxiv.org/abs/2005.08139

  • 代码:https://github.com/cxy1997/3D_adapt_auto_driving

MLCVNet: Multi-Level Context VoteNet for 3D Object Detection

  • 论文:https://arxiv.org/abs/2004.05679
  • 代码:https://github.com/NUAAXQ/MLCVNet

3DSSD: Point-based 3D Single Stage Object Detector

  • CVPR 2020 Oral

  • 论文:https://arxiv.org/abs/2002.10187

  • 代码:https://github.com/tomztyang/3DSSD

Disp R-CNN: Stereo 3D Object Detection via Shape Prior Guided Instance Disparity Estimation

  • 论文:https://arxiv.org/abs/2004.03572

  • 代码:https://github.com/zju3dv/disprcn

End-to-End Pseudo-LiDAR for Image-Based 3D Object Detection

  • 论文:https://arxiv.org/abs/2004.03080

  • 代码:https://github.com/mileyan/pseudo-LiDAR_e2e

DSGN: Deep Stereo Geometry Network for 3D Object Detection

  • 论文:https://arxiv.org/abs/2001.03398
  • 代码:https://github.com/chenyilun95/DSGN

LiDAR-based Online 3D Video Object Detection with Graph-based Message Passing and Spatiotemporal Transformer Attention

  • 论文:https://arxiv.org/abs/2004.01389
  • 代码:https://github.com/yinjunbo/3DVID

PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection

  • 论文:https://arxiv.org/abs/1912.13192

  • 代码:https://github.com/sshaoshuai/PV-RCNN

Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud

  • 论文:https://arxiv.org/abs/2003.01251
  • 代码:https://github.com/WeijingShi/Point-GNN

视频目标检测

Memory Enhanced Global-Local Aggregation for Video Object Detection

论文:https://arxiv.org/abs/2003.12063

代码:https://github.com/Scalsol/mega.pytorch

目标跟踪

SiamCAR: Siamese Fully Convolutional Classification and Regression for Visual Tracking

  • 论文:https://arxiv.org/abs/1911.07241
  • 代码:https://github.com/ohhhyeahhh/SiamCAR

D3S – A Discriminative Single Shot Segmentation Tracker

  • 论文:https://arxiv.org/abs/1911.08862
  • 代码:https://github.com/alanlukezic/d3s

ROAM: Recurrently Optimizing Tracking Model

  • 论文:https://arxiv.org/abs/1907.12006

  • 代码:https://github.com/skyoung/ROAM

Siam R-CNN: Visual Tracking by Re-Detection

  • 主页:https://www.vision.rwth-aachen.de/page/siamrcnn
  • 论文:https://arxiv.org/abs/1911.12836
  • 论文2:https://www.vision.rwth-aachen.de/media/papers/192/siamrcnn.pdf
  • 代码:https://github.com/VisualComputingInstitute/SiamR-CNN

Cooling-Shrinking Attack: Blinding the Tracker with Imperceptible Noises

  • 论文:https://arxiv.org/abs/2003.09595
  • 代码:https://github.com/MasterBin-IIAU/CSA

High-Performance Long-Term Tracking with Meta-Updater

  • 论文:https://arxiv.org/abs/2004.00305

  • 代码:https://github.com/Daikenan/LTMU

AutoTrack: Towards High-Performance Visual Tracking for UAV with Automatic Spatio-Temporal Regularization

  • 论文:https://arxiv.org/abs/2003.12949

  • 代码:https://github.com/vision4robotics/AutoTrack

Probabilistic Regression for Visual Tracking

  • 论文:https://arxiv.org/abs/2003.12565
  • 代码:https://github.com/visionml/pytracking

MAST: A Memory-Augmented Self-supervised Tracker

  • 论文:https://arxiv.org/abs/2002.07793
  • 代码:https://github.com/zlai0/MAST

Siamese Box Adaptive Network for Visual Tracking

  • 论文:https://arxiv.org/abs/2003.06761
  • 代码:https://github.com/hqucv/siamban

多目标跟踪

3D-ZeF: A 3D Zebrafish Tracking Benchmark Dataset

  • 主页:https://vap.aau.dk/3d-zef/
  • 论文:https://arxiv.org/abs/2006.08466
  • 代码:https://bitbucket.org/aauvap/3d-zef/src/master/
  • 数据集:https://motchallenge.net/data/3D-ZeF20

语义分割

FDA: Fourier Domain Adaptation for Semantic Segmentation

  • 论文:https://arxiv.org/abs/2004.05498

  • 代码:https://github.com/YanchaoYang/FDA

Super-BPD: Super Boundary-to-Pixel Direction for Fast Image Segmentation

  • 论文:暂无

  • 代码:https://github.com/JianqiangWan/Super-BPD

Single-Stage Semantic Segmentation from Image Labels

  • 论文:https://arxiv.org/abs/2005.08104

  • 代码:https://github.com/visinf/1-stage-wseg

Learning Texture Invariant Representation for Domain Adaptation of Semantic Segmentation

  • 论文:https://arxiv.org/abs/2003.00867
  • 代码:https://github.com/MyeongJin-Kim/Learning-Texture-Invariant-Representation

MSeg: A Composite Dataset for Multi-domain Semantic Segmentation

  • 论文:http://vladlen.info/papers/MSeg.pdf
  • 代码:https://github.com/mseg-dataset/mseg-api

CascadePSP: Toward Class-Agnostic and Very High-Resolution Segmentation via Global and Local Refinement

  • 论文:https://arxiv.org/abs/2005.02551
  • 代码:https://github.com/hkchengrex/CascadePSP

Unsupervised Intra-domain Adaptation for Semantic Segmentation through Self-Supervision

  • Oral
  • 论文:https://arxiv.org/abs/2004.07703
  • 代码:https://github.com/feipan664/IntraDA

Self-supervised Equivariant Attention Mechanism for Weakly Supervised Semantic Segmentation

  • 论文:https://arxiv.org/abs/2004.04581
  • 代码:https://github.com/YudeWang/SEAM

Temporally Distributed Networks for Fast Video Segmentation

  • 论文:https://arxiv.org/abs/2004.01800

  • 代码:https://github.com/feinanshan/TDNet

Context Prior for Scene Segmentation

  • 论文:https://arxiv.org/abs/2004.01547

  • 代码:https://git.io/ContextPrior

Strip Pooling: Rethinking Spatial Pooling for Scene Parsing

  • 论文:https://arxiv.org/abs/2003.13328

  • 代码:https://github.com/Andrew-Qibin/SPNet

Cars Can’t Fly up in the Sky: Improving Urban-Scene Segmentation via Height-driven Attention Networks

  • 论文:https://arxiv.org/abs/2003.05128
  • 代码:https://github.com/shachoi/HANet

Learning Dynamic Routing for Semantic Segmentation

  • 论文:https://arxiv.org/abs/2003.10401

  • 代码:https://github.com/yanwei-li/DynamicRouting

实例分割

D2Det: Towards High Quality Object Detection and Instance Segmentation

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/Cao_D2Det_Towards_High_Quality_Object_Detection_and_Instance_Segmentation_CVPR_2020_paper.pdf

  • 代码:https://github.com/JialeCao001/D2Det

PolarMask: Single Shot Instance Segmentation with Polar Representation

  • 论文:https://arxiv.org/abs/1909.13226
  • 代码:https://github.com/xieenze/PolarMask
  • 解读:https://zhuanlan.zhihu.com/p/84890413

CenterMask : Real-Time Anchor-Free Instance Segmentation

  • 论文:https://arxiv.org/abs/1911.06667
  • 代码:https://github.com/youngwanLEE/CenterMask

BlendMask: Top-Down Meets Bottom-Up for Instance Segmentation

  • 论文:https://arxiv.org/abs/2001.00309
  • 代码:https://github.com/aim-uofa/AdelaiDet

Deep Snake for Real-Time Instance Segmentation

  • 论文:https://arxiv.org/abs/2001.01629
  • 代码:https://github.com/zju3dv/snake

Mask Encoding for Single Shot Instance Segmentation

  • 论文:https://arxiv.org/abs/2003.11712

  • 代码:https://github.com/aim-uofa/AdelaiDet

全景分割

Video Panoptic Segmentation

  • 论文:https://arxiv.org/abs/2006.11339
  • 代码:https://github.com/mcahny/vps
  • 数据集:https://www.dropbox.com/s/ecem4kq0fdkver4/cityscapes-vps-dataset-1.0.zip?dl=0

Pixel Consensus Voting for Panoptic Segmentation

  • 论文:https://arxiv.org/abs/2004.01849
  • 代码:还未公布

BANet: Bidirectional Aggregation Network with Occlusion Handling for Panoptic Segmentation

论文:https://arxiv.org/abs/2003.14031

代码:https://github.com/Mooonside/BANet

视频目标分割

A Transductive Approach for Video Object Segmentation

  • 论文:https://arxiv.org/abs/2004.07193

  • 代码:https://github.com/microsoft/transductive-vos.pytorch

State-Aware Tracker for Real-Time Video Object Segmentation

  • 论文:https://arxiv.org/abs/2003.00482

  • 代码:https://github.com/MegviiDetection/video_analyst

Learning Fast and Robust Target Models for Video Object Segmentation

  • 论文:https://arxiv.org/abs/2003.00908
  • 代码:https://github.com/andr345/frtm-vos

Learning Video Object Segmentation from Unlabeled Videos

  • 论文:https://arxiv.org/abs/2003.05020
  • 代码:https://github.com/carrierlxk/MuG

超像素分割

Superpixel Segmentation with Fully Convolutional Networks

  • 论文:https://arxiv.org/abs/2003.12929
  • 代码:https://github.com/fuy34/superpixel_fcn

交互式图像分割

Interactive Object Segmentation with Inside-Outside Guidance

  • 论文下载链接:http://openaccess.thecvf.com/content_CVPR_2020/papers/Zhang_Interactive_Object_Segmentation_With_Inside-Outside_Guidance_CVPR_2020_paper.pdf
  • 代码:https://github.com/shiyinzhang/Inside-Outside-Guidance
  • 数据集:https://github.com/shiyinzhang/Pixel-ImageNet

NAS

AOWS: Adaptive and optimal network width search with latency constraints

  • 论文:https://arxiv.org/abs/2005.10481
  • 代码:https://github.com/bermanmaxim/AOWS

Densely Connected Search Space for More Flexible Neural Architecture Search

  • 论文:https://arxiv.org/abs/1906.09607

  • 代码:https://github.com/JaminFong/DenseNAS

MTL-NAS: Task-Agnostic Neural Architecture Search towards General-Purpose Multi-Task Learning

  • 论文:https://arxiv.org/abs/2003.14058

  • 代码:https://github.com/bhpfelix/MTLNAS

FBNetV2: Differentiable Neural Architecture Search for Spatial and Channel Dimensions

  • 论文下载链接:https://arxiv.org/abs/2004.05565

  • 代码:https://github.com/facebookresearch/mobile-vision

Neural Architecture Search for Lightweight Non-Local Networks

  • 论文:https://arxiv.org/abs/2004.01961
  • 代码:https://github.com/LiYingwei/AutoNL

Rethinking Performance Estimation in Neural Architecture Search

  • 论文:https://arxiv.org/abs/2005.09917
  • 代码:https://github.com/zhengxiawu/rethinking_performance_estimation_in_NAS
  • 解读1:https://www.zhihu.com/question/372070853/answer/1035234510
  • 解读2:https://zhuanlan.zhihu.com/p/111167409

CARS: Continuous Evolution for Efficient Neural Architecture Search

  • 论文:https://arxiv.org/abs/1909.04977
  • 代码(即将开源):https://github.com/huawei-noah/CARS

GAN

SEAN: Image Synthesis with Semantic Region-Adaptive Normalization

  • 论文:https://arxiv.org/abs/1911.12861
  • 代码:https://github.com/ZPdesu/SEAN

Reusing Discriminators for Encoding: Towards Unsupervised Image-to-Image Translation

  • 论文地址:http://openaccess.thecvf.com/content_CVPR_2020/html/Chen_Reusing_Discriminators_for_Encoding_Towards_Unsupervised_Image-to-Image_Translation_CVPR_2020_paper.html
  • 代码地址:https://github.com/alpc91/NICE-GAN-pytorch

Distribution-induced Bidirectional Generative Adversarial Network for Graph Representation Learning

  • 论文:https://arxiv.org/abs/1912.01899
  • 代码:https://github.com/SsGood/DBGAN

PSGAN: Pose and Expression Robust Spatial-Aware GAN for Customizable Makeup Transfer

  • 论文:https://arxiv.org/abs/1909.06956
  • 代码:https://github.com/wtjiang98/PSGAN

Semantically Mutil-modal Image Synthesis

  • 主页:http://seanseattle.github.io/SMIS
  • 论文:https://arxiv.org/abs/2003.12697
  • 代码:https://github.com/Seanseattle/SMIS

Unpaired Portrait Drawing Generation via Asymmetric Cycle Mapping

  • 论文:https://yiranran.github.io/files/CVPR2020_Unpaired%20Portrait%20Drawing%20Generation%20via%20Asymmetric%20Cycle%20Mapping.pdf
  • 代码:https://github.com/yiranran/Unpaired-Portrait-Drawing

Learning to Cartoonize Using White-box Cartoon Representations

  • 论文:https://github.com/SystemErrorWang/White-box-Cartoonization/blob/master/paper/06791.pdf

  • 主页:https://systemerrorwang.github.io/White-box-Cartoonization/

  • 代码:https://github.com/SystemErrorWang/White-box-Cartoonization

  • 解读:https://zhuanlan.zhihu.com/p/117422157

  • Demo视频:https://www.bilibili.com/video/av56708333

GAN Compression: Efficient Architectures for Interactive Conditional GANs

  • 论文:https://arxiv.org/abs/2003.08936

  • 代码:https://github.com/mit-han-lab/gan-compression

Watch your Up-Convolution: CNN Based Generative Deep Neural Networks are Failing to Reproduce Spectral Distributions

  • 论文:https://arxiv.org/abs/2003.01826
  • 代码:https://github.com/cc-hpc-itwm/UpConv

Re-ID

High-Order Information Matters: Learning Relation and Topology for Occluded Person Re-Identification

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/html/Wang_High-Order_Information_Matters_Learning_Relation_and_Topology_for_Occluded_Person_CVPR_2020_paper.html
  • 代码:https://github.com/wangguanan/HOReID

COCAS: A Large-Scale Clothes Changing Person Dataset for Re-identification

  • 论文:https://arxiv.org/abs/2005.07862

  • 数据集:暂无

Transferable, Controllable, and Inconspicuous Adversarial Attacks on Person Re-identification With Deep Mis-Ranking

  • 论文:https://arxiv.org/abs/2004.04199

  • 代码:https://github.com/whj363636/Adversarial-attack-on-Person-ReID-With-Deep-Mis-Ranking

Pose-guided Visible Part Matching for Occluded Person ReID

  • 论文:https://arxiv.org/abs/2004.00230
  • 代码:https://github.com/hh23333/PVPM

Weakly supervised discriminative feature learning with state information for person identification

  • 论文:https://arxiv.org/abs/2002.11939
  • 代码:https://github.com/KovenYu/state-information

3D点云(分类/分割/配准等)

3D点云卷积

PointASNL: Robust Point Clouds Processing using Nonlocal Neural Networks with Adaptive Sampling

  • 论文:https://arxiv.org/abs/2003.00492
  • 代码:https://github.com/yanx27/PointASNL

Global-Local Bidirectional Reasoning for Unsupervised Representation Learning of 3D Point Clouds

  • 论文下载链接:https://arxiv.org/abs/2003.12971

  • 代码:https://github.com/raoyongming/PointGLR

Grid-GCN for Fast and Scalable Point Cloud Learning

  • 论文:https://arxiv.org/abs/1912.02984

  • 代码:https://github.com/Xharlie/Grid-GCN

FPConv: Learning Local Flattening for Point Convolution

  • 论文:https://arxiv.org/abs/2002.10701
  • 代码:https://github.com/lyqun/FPConv

3D点云分类

PointAugment: an Auto-Augmentation Framework for Point Cloud Classification

  • 论文:https://arxiv.org/abs/2002.10876
  • 代码(即将开源): https://github.com/liruihui/PointAugment/

3D点云语义分割

RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds

  • 论文:https://arxiv.org/abs/1911.11236

  • 代码:https://github.com/QingyongHu/RandLA-Net

  • 解读:https://zhuanlan.zhihu.com/p/105433460

Weakly Supervised Semantic Point Cloud Segmentation:Towards 10X Fewer Labels

  • 论文:https://arxiv.org/abs/2004.04091

  • 代码:https://github.com/alex-xun-xu/WeakSupPointCloudSeg

PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation

  • 论文:https://arxiv.org/abs/2003.14032
  • 代码:https://github.com/edwardzhou130/PolarSeg

Learning to Segment 3D Point Clouds in 2D Image Space

  • 论文:https://arxiv.org/abs/2003.05593

  • 代码:https://github.com/WPI-VISLab/Learning-to-Segment-3D-Point-Clouds-in-2D-Image-Space

3D点云实例分割

PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation

  • 论文:https://arxiv.org/abs/2004.01658
  • 代码:https://github.com/Jia-Research-Lab/PointGroup

3D点云配准

Feature-metric Registration: A Fast Semi-supervised Approach for Robust Point Cloud Registration without Correspondences

  • 论文:https://arxiv.org/abs/2005.01014
  • 代码:https://github.com/XiaoshuiHuang/fmr

D3Feat: Joint Learning of Dense Detection and Description of 3D Local Features

  • 论文:https://arxiv.org/abs/2003.03164
  • 代码:https://github.com/XuyangBai/D3Feat

RPM-Net: Robust Point Matching using Learned Features

  • 论文:https://arxiv.org/abs/2003.13479
  • 代码:https://github.com/yewzijian/RPMNet

3D点云补全

Cascaded Refinement Network for Point Cloud Completion

  • 论文:https://arxiv.org/abs/2004.03327
  • 代码:https://github.com/xiaogangw/cascaded-point-completion

3D点云目标跟踪

P2B: Point-to-Box Network for 3D Object Tracking in Point Clouds

  • 论文:https://arxiv.org/abs/2005.13888
  • 代码:https://github.com/HaozheQi/P2B

其他

An Efficient PointLSTM for Point Clouds Based Gesture Recognition

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/html/Min_An_Efficient_PointLSTM_for_Point_Clouds_Based_Gesture_Recognition_CVPR_2020_paper.html
  • 代码:https://github.com/Blueprintf/pointlstm-gesture-recognition-pytorch

人脸

人脸识别

CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition

  • 论文:https://arxiv.org/abs/2004.00288

  • 代码:https://github.com/HuangYG123/CurricularFace

Learning Meta Face Recognition in Unseen Domains

  • 论文:https://arxiv.org/abs/2003.07733
  • 代码:https://github.com/cleardusk/MFR
  • 解读:https://mp.weixin.qq.com/s/YZoEnjpnlvb90qSI3xdJqQ

人脸检测

人脸活体检测

Searching Central Difference Convolutional Networks for Face Anti-Spoofing

  • 论文:https://arxiv.org/abs/2003.04092

  • 代码:https://github.com/ZitongYu/CDCN

人脸表情识别

Suppressing Uncertainties for Large-Scale Facial Expression Recognition

  • 论文:https://arxiv.org/abs/2002.10392

  • 代码(即将开源):https://github.com/kaiwang960112/Self-Cure-Network

人脸转正

Rotate-and-Render: Unsupervised Photorealistic Face Rotation from Single-View Images

  • 论文:https://arxiv.org/abs/2003.08124
  • 代码:https://github.com/Hangz-nju-cuhk/Rotate-and-Render

人脸3D重建

AvatarMe: Realistically Renderable 3D Facial Reconstruction "in-the-wild"

  • 论文:https://arxiv.org/abs/2003.13845
  • 数据集:https://github.com/lattas/AvatarMe

FaceScape: a Large-scale High Quality 3D Face Dataset and Detailed Riggable 3D Face Prediction

  • 论文:https://arxiv.org/abs/2003.13989
  • 代码:https://github.com/zhuhao-nju/facescape

人体姿态估计(2D/3D)

2D人体姿态估计

TransMoMo: Invariance-Driven Unsupervised Video Motion Retargeting

  • 主页:https://yzhq97.github.io/transmomo/

  • 论文:https://arxiv.org/abs/2003.14401

  • 代码:https://github.com/yzhq97/transmomo.pytorch

HigherHRNet: Scale-Aware Representation Learning for Bottom-Up Human Pose Estimation

  • 论文:https://arxiv.org/abs/1908.10357
  • 代码:https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation

The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation

  • 论文:https://arxiv.org/abs/1911.07524
  • 代码:https://github.com/HuangJunJie2017/UDP-Pose
  • 解读:https://zhuanlan.zhihu.com/p/92525039

Distribution-Aware Coordinate Representation for Human Pose Estimation

  • 主页:https://ilovepose.github.io/coco/

  • 论文:https://arxiv.org/abs/1910.06278

  • 代码:https://github.com/ilovepose/DarkPose

3D人体姿态估计

Cascaded Deep Monocular 3D Human Pose Estimation With Evolutionary Training Data

  • 论文:https://arxiv.org/abs/2006.07778
  • 代码:https://github.com/Nicholasli1995/EvoSkeleton

Fusing Wearable IMUs with Multi-View Images for Human Pose Estimation: A Geometric Approach

  • 主页:https://www.zhe-zhang.com/cvpr2020

  • 论文:https://arxiv.org/abs/2003.11163

  • 代码:https://github.com/CHUNYUWANG/imu-human-pose-pytorch

Bodies at Rest: 3D Human Pose and Shape Estimation from a Pressure Image using Synthetic Data

  • 论文下载链接:https://arxiv.org/abs/2004.01166

  • 代码:https://github.com/Healthcare-Robotics/bodies-at-rest

  • 数据集:https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/KOA4ML

Self-Supervised 3D Human Pose Estimation via Part Guided Novel Image Synthesis

  • 主页:http://val.cds.iisc.ac.in/pgp-human/
  • 论文:https://arxiv.org/abs/2004.04400

Compressed Volumetric Heatmaps for Multi-Person 3D Pose Estimation

  • 论文:https://arxiv.org/abs/2004.00329
  • 代码:https://github.com/fabbrimatteo/LoCO

VIBE: Video Inference for Human Body Pose and Shape Estimation

  • 论文:https://arxiv.org/abs/1912.05656
  • 代码:https://github.com/mkocabas/VIBE

Back to the Future: Joint Aware Temporal Deep Learning 3D Human Pose Estimation

  • 论文:https://arxiv.org/abs/2002.11251
  • 代码:https://github.com/vnmr/JointVideoPose3D

Cross-View Tracking for Multi-Human 3D Pose Estimation at over 100 FPS

  • 论文:https://arxiv.org/abs/2003.03972
  • 数据集:暂无

人体解析

Correlating Edge, Pose with Parsing

  • 论文:https://arxiv.org/abs/2005.01431

  • 代码:https://github.com/ziwei-zh/CorrPM

场景文本检测

STEFANN: Scene Text Editor using Font Adaptive Neural Network

  • 主页:https://prasunroy.github.io/stefann/

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/html/Roy_STEFANN_Scene_Text_Editor_Using_Font_Adaptive_Neural_Network_CVPR_2020_paper.html

  • 代码:https://github.com/prasunroy/stefann

  • 数据集:https://drive.google.com/open?id=1sEDiX_jORh2X-HSzUnjIyZr-G9LJIw1k

ContourNet: Taking a Further Step Toward Accurate Arbitrary-Shaped Scene Text Detection

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/Wang_ContourNet_Taking_a_Further_Step_Toward_Accurate_Arbitrary-Shaped_Scene_Text_CVPR_2020_paper.pdf
  • 代码:https://github.com/wangyuxin87/ContourNet

UnrealText: Synthesizing Realistic Scene Text Images from the Unreal World

  • 论文:https://arxiv.org/abs/2003.10608
  • 代码和数据集:https://github.com/Jyouhou/UnrealText/

ABCNet: Real-time Scene Text Spotting with Adaptive Bezier-Curve Network

  • 论文:https://arxiv.org/abs/2002.10200
  • 代码(即将开源):https://github.com/Yuliang-Liu/bezier_curve_text_spotting
  • 代码(即将开源):https://github.com/aim-uofa/adet

Deep Relational Reasoning Graph Network for Arbitrary Shape Text Detection

  • 论文:https://arxiv.org/abs/2003.07493

  • 代码:https://github.com/GXYM/DRRG

场景文本识别

SEED: Semantics Enhanced Encoder-Decoder Framework for Scene Text Recognition

  • 论文:https://arxiv.org/abs/2005.10977
  • 代码:https://github.com/Pay20Y/SEED

UnrealText: Synthesizing Realistic Scene Text Images from the Unreal World

  • 论文:https://arxiv.org/abs/2003.10608
  • 代码和数据集:https://github.com/Jyouhou/UnrealText/

ABCNet: Real-time Scene Text Spotting with Adaptive Bezier-Curve Network

  • 论文:https://arxiv.org/abs/2002.10200
  • 代码(即将开源):https://github.com/aim-uofa/adet

Learn to Augment: Joint Data Augmentation and Network Optimization for Text Recognition

  • 论文:https://arxiv.org/abs/2003.06606

  • 代码:https://github.com/Canjie-Luo/Text-Image-Augmentation

特征(点)检测和描述

SuperGlue: Learning Feature Matching with Graph Neural Networks

  • 论文:https://arxiv.org/abs/1911.11763
  • 代码:https://github.com/magicleap/SuperGluePretrainedNetwork

超分辨率

图像超分辨率

Closed-Loop Matters: Dual Regression Networks for Single Image Super-Resolution

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/html/Guo_Closed-Loop_Matters_Dual_Regression_Networks_for_Single_Image_Super-Resolution_CVPR_2020_paper.html
  • 代码:https://github.com/guoyongcs/DRN

Learning Texture Transformer Network for Image Super-Resolution

  • 论文:https://arxiv.org/abs/2006.04139

  • 代码:https://github.com/FuzhiYang/TTSR

Image Super-Resolution with Cross-Scale Non-Local Attention and Exhaustive Self-Exemplars Mining

  • 论文:https://arxiv.org/abs/2006.01424
  • 代码:https://github.com/SHI-Labs/Cross-Scale-Non-Local-Attention

Structure-Preserving Super Resolution with Gradient Guidance

  • 论文:https://arxiv.org/abs/2003.13081

  • 代码:https://github.com/Maclory/SPSR

Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Analysis and a New Strategy

论文:https://arxiv.org/abs/2004.00448

代码:https://github.com/clovaai/cutblur

视频超分辨率

TDAN: Temporally-Deformable Alignment Network for Video Super-Resolution

  • 论文:https://arxiv.org/abs/1812.02898
  • 代码:https://github.com/YapengTian/TDAN-VSR-CVPR-2020

Space-Time-Aware Multi-Resolution Video Enhancement

  • 主页:https://alterzero.github.io/projects/STAR.html
  • 论文:http://arxiv.org/abs/2003.13170
  • 代码:https://github.com/alterzero/STARnet

Zooming Slow-Mo: Fast and Accurate One-Stage Space-Time Video Super-Resolution

  • 论文:https://arxiv.org/abs/2002.11616
  • 代码:https://github.com/Mukosame/Zooming-Slow-Mo-CVPR-2020

模型压缩/剪枝

DMCP: Differentiable Markov Channel Pruning for Neural Networks

  • 论文:https://arxiv.org/abs/2005.03354
  • 代码:https://github.com/zx55/dmcp

Forward and Backward Information Retention for Accurate Binary Neural Networks

  • 论文:https://arxiv.org/abs/1909.10788

  • 代码:https://github.com/htqin/IR-Net

Towards Efficient Model Compression via Learned Global Ranking

  • 论文:https://arxiv.org/abs/1904.12368
  • 代码:https://github.com/cmu-enyac/LeGR

HRank: Filter Pruning using High-Rank Feature Map

  • 论文:http://arxiv.org/abs/2002.10179
  • 代码:https://github.com/lmbxmu/HRank

GAN Compression: Efficient Architectures for Interactive Conditional GANs

  • 论文:https://arxiv.org/abs/2003.08936

  • 代码:https://github.com/mit-han-lab/gan-compression

Group Sparsity: The Hinge Between Filter Pruning and Decomposition for Network Compression

  • 论文:https://arxiv.org/abs/2003.08935

  • 代码:https://github.com/ofsoundof/group_sparsity

视频理解/行为识别

Oops! Predicting Unintentional Action in Video

  • 主页:https://oops.cs.columbia.edu/

  • 论文:https://arxiv.org/abs/1911.11206

  • 代码:https://github.com/cvlab-columbia/oops

  • 数据集:https://oops.cs.columbia.edu/data

PREDICT & CLUSTER: Unsupervised Skeleton Based Action Recognition

  • 论文:https://arxiv.org/abs/1911.12409
  • 代码:https://github.com/shlizee/Predict-Cluster

Intra- and Inter-Action Understanding via Temporal Action Parsing

  • 论文:https://arxiv.org/abs/2005.10229
  • 主页和数据集:https://sdolivia.github.io/TAPOS/

3DV: 3D Dynamic Voxel for Action Recognition in Depth Video

  • 论文:https://arxiv.org/abs/2005.05501
  • 代码:https://github.com/3huo/3DV-Action

FineGym: A Hierarchical Video Dataset for Fine-grained Action Understanding

  • 主页:https://sdolivia.github.io/FineGym/
  • 论文:https://arxiv.org/abs/2004.06704

TEA: Temporal Excitation and Aggregation for Action Recognition

  • 论文:https://arxiv.org/abs/2004.01398

  • 代码:https://github.com/Phoenix1327/tea-action-recognition

X3D: Expanding Architectures for Efficient Video Recognition

  • 论文:https://arxiv.org/abs/2004.04730

  • 代码:https://github.com/facebookresearch/SlowFast

Temporal Pyramid Network for Action Recognition

  • 主页:https://decisionforce.github.io/TPN

  • 论文:https://arxiv.org/abs/2004.03548

  • 代码:https://github.com/decisionforce/TPN

基于骨架的动作识别

Disentangling and Unifying Graph Convolutions for Skeleton-Based Action Recognition

  • 论文:https://arxiv.org/abs/2003.14111
  • 代码:https://github.com/kenziyuliu/ms-g3d

人群计数

深度估计

BiFuse: Monocular 360◦ Depth Estimation via Bi-Projection Fusion

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/Wang_BiFuse_Monocular_360_Depth_Estimation_via_Bi-Projection_Fusion_CVPR_2020_paper.pdf
  • 代码:https://github.com/Yeh-yu-hsuan/BiFuse

Focus on defocus: bridging the synthetic to real domain gap for depth estimation

  • 论文:https://arxiv.org/abs/2005.09623
  • 代码:https://github.com/dvl-tum/defocus-net

Bi3D: Stereo Depth Estimation via Binary Classifications

  • 论文:https://arxiv.org/abs/2005.07274

  • 代码:https://github.com/NVlabs/Bi3D

AANet: Adaptive Aggregation Network for Efficient Stereo Matching

  • 论文:https://arxiv.org/abs/2004.09548
  • 代码:https://github.com/haofeixu/aanet

Towards Better Generalization: Joint Depth-Pose Learning without PoseNet

  • 论文:https://github.com/B1ueber2y/TrianFlow

  • 代码:https://github.com/B1ueber2y/TrianFlow

单目深度估计

On the uncertainty of self-supervised monocular depth estimation

  • 论文:https://arxiv.org/abs/2005.06209
  • 代码:https://github.com/mattpoggi/mono-uncertainty

3D Packing for Self-Supervised Monocular Depth Estimation

  • 论文:https://arxiv.org/abs/1905.02693
  • 代码:https://github.com/TRI-ML/packnet-sfm
  • Demo视频:https://www.bilibili.com/video/av70562892/

Domain Decluttering: Simplifying Images to Mitigate Synthetic-Real Domain Shift and Improve Depth Estimation

  • 论文:https://arxiv.org/abs/2002.12114
  • 代码:https://github.com/yzhao520/ARC

6D目标姿态估计

PVN3D: A Deep Point-wise 3D Keypoints Voting Network for 6DoF Pose Estimation

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/He_PVN3D_A_Deep_Point-Wise_3D_Keypoints_Voting_Network_for_6DoF_CVPR_2020_paper.pdf
  • 代码:https://github.com/ethnhe/PVN3D

MoreFusion: Multi-object Reasoning for 6D Pose Estimation from Volumetric Fusion

  • 论文:https://arxiv.org/abs/2004.04336
  • 代码:https://github.com/wkentaro/morefusion

EPOS: Estimating 6D Pose of Objects with Symmetries

主页:http://cmp.felk.cvut.cz/epos

论文:https://arxiv.org/abs/2004.00605

G2L-Net: Global to Local Network for Real-time 6D Pose Estimation with Embedding Vector Features

  • 论文:https://arxiv.org/abs/2003.11089

  • 代码:https://github.com/DC1991/G2L_Net

手势估计

HOPE-Net: A Graph-based Model for Hand-Object Pose Estimation

  • 论文:https://arxiv.org/abs/2004.00060

  • 主页:http://vision.sice.indiana.edu/projects/hopenet

Monocular Real-time Hand Shape and Motion Capture using Multi-modal Data

  • 论文:https://arxiv.org/abs/2003.09572

  • 代码:https://github.com/CalciferZh/minimal-hand

显著性检测

JL-DCF: Joint Learning and Densely-Cooperative Fusion Framework for RGB-D Salient Object Detection

  • 论文:https://arxiv.org/abs/2004.08515

  • 代码:https://github.com/kerenfu/JLDCF/

UC-Net: Uncertainty Inspired RGB-D Saliency Detection via Conditional Variational Autoencoders

  • 主页:http://dpfan.net/d3netbenchmark/

  • 论文:https://arxiv.org/abs/2004.05763

  • 代码:https://github.com/JingZhang617/UCNet

去噪

A Physics-based Noise Formation Model for Extreme Low-light Raw Denoising

  • 论文:https://arxiv.org/abs/2003.12751

  • 代码:https://github.com/Vandermode/NoiseModel

CycleISP: Real Image Restoration via Improved Data Synthesis

  • 论文:https://arxiv.org/abs/2003.07761

  • 代码:https://github.com/swz30/CycleISP

去雨

Multi-Scale Progressive Fusion Network for Single Image Deraining

  • 论文:https://arxiv.org/abs/2003.10985
  • 代码:https://github.com/kuihua/MSPFN

Detail-recovery Image Deraining via Context Aggregation Networks

  • 论文:https://openaccess.thecvf.com/content_CVPR_2020/html/Deng_Detail-recovery_Image_Deraining_via_Context_Aggregation_Networks_CVPR_2020_paper.html
  • 代码:https://github.com/Dengsgithub/DRD-Net

去模糊

视频去模糊

Cascaded Deep Video Deblurring Using Temporal Sharpness Prior

  • 主页:https://csbhr.github.io/projects/cdvd-tsp/index.html
  • 论文:https://arxiv.org/abs/2004.02501
  • 代码:https://github.com/csbhr/CDVD-TSP

去雾

Domain Adaptation for Image Dehazing

  • 论文:https://arxiv.org/abs/2005.04668

  • 代码:https://github.com/HUSTSYJ/DA_dahazing

Multi-Scale Boosted Dehazing Network with Dense Feature Fusion

  • 论文:https://arxiv.org/abs/2004.13388

  • 代码:https://github.com/BookerDeWitt/MSBDN-DFF

特征点检测与描述

ASLFeat: Learning Local Features of Accurate Shape and Localization

  • 论文:https://arxiv.org/abs/2003.10071

  • 代码:https://github.com/lzx551402/aslfeat

视觉问答(VQA)

VC R-CNN:Visual Commonsense R-CNN

  • 论文:https://arxiv.org/abs/2002.12204
  • 代码:https://github.com/Wangt-CN/VC-R-CNN

视频问答(VideoQA)

Hierarchical Conditional Relation Networks for Video Question Answering

  • 论文:https://arxiv.org/abs/2002.10698
  • 代码:https://github.com/thaolmk54/hcrn-videoqa

视觉语言导航

Towards Learning a Generic Agent for Vision-and-Language Navigation via Pre-training

  • 论文:https://arxiv.org/abs/2002.10638
  • 代码(即将开源):https://github.com/weituo12321/PREVALENT

视频压缩

Learning for Video Compression with Hierarchical Quality and Recurrent Enhancement

  • 论文:https://arxiv.org/abs/2003.01966
  • 代码:https://github.com/RenYang-home/HLVC

视频插帧

AdaCoF: Adaptive Collaboration of Flows for Video Frame Interpolation

  • 论文:https://arxiv.org/abs/1907.10244
  • 代码:https://github.com/HyeongminLEE/AdaCoF-pytorch

FeatureFlow: Robust Video Interpolation via Structure-to-Texture Generation

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/html/Gui_FeatureFlow_Robust_Video_Interpolation_via_Structure-to-Texture_Generation_CVPR_2020_paper.html

  • 代码:https://github.com/CM-BF/FeatureFlow

Zooming Slow-Mo: Fast and Accurate One-Stage Space-Time Video Super-Resolution

  • 论文:https://arxiv.org/abs/2002.11616
  • 代码:https://github.com/Mukosame/Zooming-Slow-Mo-CVPR-2020

Space-Time-Aware Multi-Resolution Video Enhancement

  • 主页:https://alterzero.github.io/projects/STAR.html
  • 论文:http://arxiv.org/abs/2003.13170
  • 代码:https://github.com/alterzero/STARnet

Scene-Adaptive Video Frame Interpolation via Meta-Learning

  • 论文:https://arxiv.org/abs/2004.00779
  • 代码:https://github.com/myungsub/meta-interpolation

Softmax Splatting for Video Frame Interpolation

  • 主页:http://sniklaus.com/papers/softsplat
  • 论文:https://arxiv.org/abs/2003.05534
  • 代码:https://github.com/sniklaus/softmax-splatting

风格迁移

Diversified Arbitrary Style Transfer via Deep Feature Perturbation

  • 论文:https://arxiv.org/abs/1909.08223
  • 代码:https://github.com/EndyWon/Deep-Feature-Perturbation

Collaborative Distillation for Ultra-Resolution Universal Style Transfer

  • 论文:https://arxiv.org/abs/2003.08436

  • 代码:https://github.com/mingsun-tse/collaborative-distillation

车道线检测

Inter-Region Affinity Distillation for Road Marking Segmentation

  • 论文:https://arxiv.org/abs/2004.05304
  • 代码:https://github.com/cardwing/Codes-for-IntRA-KD

"人-物"交互(HOT)检测

PPDM: Parallel Point Detection and Matching for Real-time Human-Object Interaction Detection

  • 论文:https://arxiv.org/abs/1912.12898
  • 代码:https://github.com/YueLiao/PPDM

Detailed 2D-3D Joint Representation for Human-Object Interaction

  • 论文:https://arxiv.org/abs/2004.08154

  • 代码:https://github.com/DirtyHarryLYL/DJ-RN

Cascaded Human-Object Interaction Recognition

  • 论文:https://arxiv.org/abs/2003.04262

  • 代码:https://github.com/tfzhou/C-HOI

VSGNet: Spatial Attention Network for Detecting Human Object Interactions Using Graph Convolutions

  • 论文:https://arxiv.org/abs/2003.05541
  • 代码:https://github.com/ASMIftekhar/VSGNet

轨迹预测

The Garden of Forking Paths: Towards Multi-Future Trajectory Prediction

  • 论文:https://arxiv.org/abs/1912.06445
  • 代码:https://github.com/JunweiLiang/Multiverse
  • 数据集:https://next.cs.cmu.edu/multiverse/

Social-STGCNN: A Social Spatio-Temporal Graph Convolutional Neural Network for Human Trajectory Prediction

  • 论文:https://arxiv.org/abs/2002.11927
  • 代码:https://github.com/abduallahmohamed/Social-STGCNN

运动预测

Collaborative Motion Prediction via Neural Motion Message Passing

  • 论文:https://arxiv.org/abs/2003.06594
  • 代码:https://github.com/PhyllisH/NMMP

MotionNet: Joint Perception and Motion Prediction for Autonomous Driving Based on Bird’s Eye View Maps

  • 论文:https://arxiv.org/abs/2003.06754

  • 代码:https://github.com/pxiangwu/MotionNet

光流估计

Learning by Analogy: Reliable Supervision from Transformations for Unsupervised Optical Flow Estimation

  • 论文:https://arxiv.org/abs/2003.13045
  • 代码:https://github.com/lliuz/ARFlow

图像检索

Evade Deep Image Retrieval by Stashing Private Images in the Hash Space

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/html/Xiao_Evade_Deep_Image_Retrieval_by_Stashing_Private_Images_in_the_CVPR_2020_paper.html
  • 代码:https://github.com/sugarruy/hashstash

虚拟试衣

Towards Photo-Realistic Virtual Try-On by Adaptively Generating↔Preserving Image Content

  • 论文:https://arxiv.org/abs/2003.05863
  • 代码:https://github.com/switchablenorms/DeepFashion_Try_On

HDR

Single-Image HDR Reconstruction by Learning to Reverse the Camera Pipeline

  • 主页:https://www.cmlab.csie.ntu.edu.tw/~yulunliu/SingleHDR

  • 论文下载链接:https://www.cmlab.csie.ntu.edu.tw/~yulunliu/SingleHDR_/00942.pdf

  • 代码:https://github.com/alex04072000/SingleHDR

对抗样本

Enhancing Cross-Task Black-Box Transferability of Adversarial Examples With Dispersion Reduction

  • 论文:https://openaccess.thecvf.com/content_CVPR_2020/papers/Lu_Enhancing_Cross-Task_Black-Box_Transferability_of_Adversarial_Examples_With_Dispersion_Reduction_CVPR_2020_paper.pdf
  • 代码:https://github.com/erbloo/dr_cvpr20

Towards Large yet Imperceptible Adversarial Image Perturbations with Perceptual Color Distance

  • 论文:https://arxiv.org/abs/1911.02466
  • 代码:https://github.com/ZhengyuZhao/PerC-Adversarial

三维重建

Unsupervised Learning of Probably Symmetric Deformable 3D Objects from Images in the Wild

  • CVPR 2020 Best Paper
  • 主页:https://elliottwu.com/projects/unsup3d/
  • 论文:https://arxiv.org/abs/1911.11130
  • 代码:https://github.com/elliottwu/unsup3d

Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization

  • 主页:https://shunsukesaito.github.io/PIFuHD/

  • 论文:https://arxiv.org/abs/2004.00452

  • 代码:https://github.com/facebookresearch/pifuhd

  • 论文:http://openaccess.thecvf.com/content_CVPR_2020/papers/Patel_TailorNet_Predicting_Clothing_in_3D_as_a_Function_of_Human_CVPR_2020_paper.pdf

  • 代码:https://github.com/chaitanya100100/TailorNet

  • 数据集:https://github.com/zycliao/TailorNet_dataset

Implicit Functions in Feature Space for 3D Shape Reconstruction and Completion