![](https://img-blog.csdnimg.cn/20201014180756918.png?x-oss-process=image/resize,m_fixed,h_64,w_64)
论文笔记
文章平均质量分 93
eight_Jessen
这个作者很懒,什么都没留下…
展开
-
ChatGLM论文解读
chatGLM论文解读原创 2024-01-25 10:50:52 · 1246 阅读 · 0 评论 -
论文笔记:A review on multi-label learning
这篇文章介绍了多标签分类的定义和评价指标、多标签学习的算法还有其他相关的任务。原创 2023-12-11 17:04:28 · 689 阅读 · 0 评论 -
论文解读VSR MuCAN: Multi-Correspondence Aggregation Network for Video Super-Resolution 2020 ECCV
MuCAN: Multi-Correspondence Aggregation Network for Video Super-ResolutionGitHub地址1.总结这篇文章作者主要在于突出利用多帧输入里面帧间和帧内的信息,对此作者分别提出了Temporal Multi-Correspondence Aggregation Module 和 Cross-Scale Nonlocal-Correspondence Aggregation Module,相比于以往的视频超分,这两个模块的功能我认为可原创 2021-01-06 11:08:20 · 1151 阅读 · 4 评论 -
论文笔记MEMC-Net TPAMI
MEMC-Net: Motion Estimation and Motion Compensation Driven Neural Network for Video Interpolation and Enhancement总结在传统的视频插帧中,通常会用到motion estimation(ME)和motion compensation(MC)。目前存在的基于光流的方法要么预测光流,要么预测补偿核,限制了计算的高效和插帧的准确。作者提出一个用于视频插帧的运动估计和运动补偿驱动的网络,并使用也给自适应原创 2020-12-25 10:24:14 · 1123 阅读 · 0 评论 -
2020CVPR VSR Space-Time-Aware Multi-Resolution Video Enhancement
Space-Time-Aware Multi-Resolution Video Enhancement1、总结同时做时间和空间的超分。高分辨率可以提高运动细节,高帧率有利于做运动对齐。文中的方法是在ST-SR期间生成潜在的低分辨率和高分辨率表示的模型组件可用于微调仅针对空间SR或时间SR的专用机制。作者提出了 Space-Time-Aware multi-Resolution Network STARnet。STARnet通过为ST-SR提供从LR到HR的直接连接,明确合并了在LR和HR空间中相互增强原创 2020-11-09 19:50:17 · 377 阅读 · 0 评论 -
2020ECCV VSR Video Super-Resolution with Recurrent Structure-Detail Network
Video Super-Resolution with Recurrent Structure-Detail Network1.总结作者提出的网络将输入分成了结构和细节两部分,这些部分被送入到由几个proposed two-stream struct-detail模块。另外,引入了自适应隐藏层模块,允许当前帧可以有选择地使用来自隐藏层状态的信息,可以增强对外观变化和累积错误的鲁棒性。分析比较:以前VSR的方法时通过显式的运动的那个补偿来实现,先通过计算参考帧和邻帧的光流,然后对齐做超分。但是密集光流原创 2020-11-07 20:27:17 · 347 阅读 · 0 评论 -
VSR论文笔记四|Frame-Recurrent Video Super-Resolution
Frame-Recurrent Video Super-Resolution1.摘要作者认为以往的做法是多个LR帧得到一个HR帧。这种的方法有两个主要的缺点:1)每个输入帧都经过多次处理和变形,从而增加了计算成本;2)每个输出帧都是根据输入帧进行独立估计的,从而限制了系统产生时间上一致的结果的能力作者提出的网络是使用预测出来的HR,继续预测后面的帧。由于其重复性,所提出的方法具有同化大量先前帧的能力,而不会增加计算需求。2.Introduction最新的视频超分辨率方法通过组合一批LR帧以估计原创 2020-11-06 15:53:08 · 536 阅读 · 0 评论 -
VSR论文笔记三| 2018CVPR Deep Video Super-Resolution Network Using Dynamic Upsampling Filters Without Expl
Deep Video Super-Resolution Network Using Dynamic Upsampling Filters Without Explicit Motion Compensation1.总结以往的方法依赖于运动估计和补偿。对运动估计的准确度要求高。同时最后输出的HR图像是通过CNN混合来自多个运动补偿输入LR帧得到的,最终的结果也比较模糊。作者提出一个基于每个像素局部的时空邻域产生动态上采样滤波器和残差图像的网络,一次阻止显式的运动补偿。最终HR图像的产生是通过直接对输入图原创 2020-11-04 20:07:30 · 1063 阅读 · 0 评论 -
VSR论文笔记二|Robust Video Super-Resolution with Learned Temporal Dynamics
1.总结视频超分提取帧间的信息很重要,作者提出了一个可以自适应选择优化范围的时序自适应网络,同时作者用一个空间对齐网络减少邻帧的的运动复杂性。具体来讲就是:首先有一个时序自适应网络。时序信息对视频超分很重要,以往有通过复杂的优化来解决但是引入了计算负担和时间负担,也有一些使用固定的temporal scale通过显式应用运动补偿来产生网络的输入。作者提出一个自适应时序网络,可以鲁棒应对各种运动类型并且选择优化的范围。网络的输入是经过运动补偿后对齐的LR帧,然后应用不同的时序size产生HRsize估计。原创 2020-11-03 17:54:14 · 496 阅读 · 0 评论 -
videoSR视频超分论文笔记一
1. Bidirectional Recurrent Convolutional Networks for Multi-Frame Super-Resolution以往的subpixel motions estimation只适用于小的运动,同时这种做法计算量大。作者设计的网络包括三个要点:前馈卷积模拟低分辨率帧与其高分辨率结果之间的视觉空间依赖性。循环卷积连接连续帧的隐藏层以了解时间依赖性。条件卷积连接之前时间戳的输入和和现在的隐藏层。使用MSE训练网络。数据集 25YUV2. SUPER-R原创 2020-10-29 10:39:27 · 1042 阅读 · 0 评论 -
2020CVPR超分系列二Deep Unfolding Network for Image SR+Meta-Transfer Learning ZSSR+Res FeatureAggregation
1、Deep Unfolding Network for Image Super-Resolution代码传送门1.1 总结作者认为:learning-base方法目前展现出相比传统model-base方法更好的结果。然而model-base方法可以解决的超分中一些问题,比如不同的缩放因子,模糊核,噪声水平。所以作者提出了一个利用了model-base和learning-base两者优势的网络。通过半二次分裂算法,可以得到由交替求解一个数据子问题和先验子问题组成的固定迭代次数,这部分可以由神经网原创 2020-09-29 18:10:31 · 1197 阅读 · 0 评论 -
2020CVPR超分列:UnpairedImage SR using Pseudo-Supervision+Data Augmentation SR+Closed-loop Matters
1、Unpaired Image Super-Resolution using Pseudo-Supervision根据unpaired 训练样本,做一个从LR源域x(∈X)x(\in X)x(∈X)到SR目标域y(∈Y)y(\in Y)y(∈Y)的映射。clean LR HR根据一个预先定义好的操作降采样y↓(∈Y↓)y_{\downarrow}(\in Y_{\downarrow})y↓(∈Y↓)。FXYF_{XY}FXY就是指GXY↓G_{XY_{\downarrow}}GXY↓和 UY↓原创 2020-09-28 10:06:34 · 533 阅读 · 0 评论 -
2019CVPR超分文章网络结构
1、Modulating Image Restoration with Continual Levels via Adaptive Feature Modification Layers调整模糊度,upscale重建加入AdaFM模块引导。2、Natural and Realistic Single Image Super-Resolution with Explicit Natural Manifold Discrimination3、ODE-inspired Network Design f原创 2020-09-22 10:29:56 · 578 阅读 · 0 评论 -
超分论文笔记2020CVPR视频超分:Zooming Slow-Mo- VSR with Temporal Group Attention-TDAN
Space-Time Video Super-Resolution (STVSR) 问题定义:从一个低像素低帧率恢复出高帧率高分辨率的视频。1.Zooming Slow-Mo: Fast and Accurate One-Stage Space-Time Video Super-Resolution代码链接1.1 总结之前的一些方法采用手工制作的正则化方法,并做出比较强的假设,这些方法限制了模型的容量和扩展到更多样的模式,同时计算量大。现在的一些深度学习的方法,一种直接的方法是组合对视频插针和视原创 2020-09-11 16:34:33 · 1006 阅读 · 0 评论 -
超分论文笔记之纹理迁移2019-2020CVPR:Image SRby Neural Texture Transfer -Learning Texture Transformer Network
1. Image Super-Resolution by Neural Texture Transfer代码地址1.1 总结使用RefS方法,当参考图像很相似时,超分的结果还不错。但是参考图像对超分结果影响很大,特别是当参考图像相似性比较低时,效果不佳。作者通过纹理细节,根据纹理相似性做超分的方法,让RefSR方法受参考图像的相似性影响比较少。相比以往在输入做match,作者在多个level做match,利用多尺度神经迁移,模型能够从具有语义相关性的Ref patches获益更多,在输入的ref im原创 2020-09-04 11:21:41 · 1326 阅读 · 0 评论 -
论文笔记之CVPR2019超分四:Second-order Attention Network SR-Real Scene Super-Resolution with Raw Images-DPSR
1、Second-order Attention Network for Single Image Super-Resolution代码链接1.1 总结以往的网络都是通过更深更大的结构来实现超分,但是他们忽略了中间层的特征联系,没有发挥出网络的表示作用。作者提出了一个second-order attention(SAN)网络,可以有更强的特征表示和特征关系学习。具体来说有一个second-order channel attention(SOCA) module通过二阶特征统计特性自适应调整chann原创 2020-09-03 20:30:37 · 2712 阅读 · 5 评论 -
论文笔记之CVPR2019超分三:PASSRnet-NatSR-AdaFMNet
1、Learning Parallax Attention for Stereo Image Super-Resolution(2019CVPR PASSRnet)1.1 方法输入:a stereo image pair as input and super-resolves the left image一个立体图像对作为输入和左图像超分结果?总体网络结构图如图所示:1.1.1 Residual Atrous Spatial Pyramid Pooling (ASPP) Module利用密集的像原创 2020-09-02 11:07:29 · 1012 阅读 · 0 评论 -
论文笔记之视频:SDC-Net_Video prediction using spatially-displaced convolution
AbstractLearn a motion vector and a kernel for each pixel and synthesize a pixel by applying the kernel at a displaced location in the source image, defined by the predicted motion vector.1.IntroductionVideo prediction taskAaccurately capture not only原创 2020-09-02 09:47:42 · 276 阅读 · 0 评论 -
论文笔记之超分二:IKC-MetaSR-ODEInspired
2019前超分文章记录 SRCNN-FSRCNN-ESPCN-VDCN-DRCN-RDN-LapSRN-SRDenseNet2019CVPR超分文章记录系列一:FSTRN-resLR-SRFBN-RBPN1.Blind Super-Resolution With Iterative Kernel (IKC)1.1 介绍当预先定义的模糊核与实际模糊核不同时,基于学习的方法将遭受严重的性能下降。将未知模糊核的超分问题我们定义为blind SR。这种设置比较符合实际真实情况。**所以这篇文章解决的是超分原创 2020-08-29 20:25:22 · 1983 阅读 · 0 评论 -
论文笔记之超分:FSTRN-resLR-SRFBN-RBPN
2014到2018的图片超分文章总结传送门1. Fast Spatio-Temporal Residual Network for Video Super-Resolution(FSTRN 2019 CVPR)1.1 网络结构网络分成了四部分:LR video shallow feature extraction net(LFENet)fast spatio-temporal residual blocks (FRBs)LR feature fusion and upsampling SR n原创 2020-08-28 10:54:58 · 1096 阅读 · 0 评论 -
论文笔记之视频:Video Compression through Image Interpolation
2.Related WorkImage CompressionProgressively encodes the image using a recurrent neural networkallow for variable compression rates with a single modelUse fully convolutional networks to handle arbitrary image sizesBottleneck:contains spatially redun原创 2020-08-26 09:27:24 · 494 阅读 · 0 评论 -
论文笔记之视频:Unsupervised Learning of Multi-Frame Optical Flow with Occlusions
Accurate estimate of optical flowThis can be attributed to the large degree of ambiguities inherent to this ill-posed problem which can only be resolved using prior knowledge about the appearance and motion of image sequences.Large datasets and obtainin原创 2020-08-25 19:35:19 · 331 阅读 · 0 评论 -
论文笔记之抓取:Efficient Grasping from RGBD Images Learning using a new Rectangle Representation
7DThe full 7-dimensional gripper configuration—its 3D location, 3D orientation and the gripper opening width.MethodA two-step learning algorithm to efficiently learn this representation.Describe a certain class of features that makes the inference in原创 2020-08-24 09:28:47 · 600 阅读 · 0 评论 -
论文笔记之抓取:Learning 6-DOF Grasping Interaction via Deep Geometry-aware 3D Representations
Contributions:(1) learn a 6-DOF grasping net from RGBD input;(2) We build a grasping dataset from demonstrations in virtual reality with rich sensory and interaction annotations, propose a data augmentation strategy for effective learning;(3) demonstra原创 2020-08-23 10:09:10 · 372 阅读 · 0 评论 -
论文笔记之视频:Slow and steady feature analysishigher order temporal coherence in video
AbstractCapture how the visual content changesGeneralize slow feature analysis to “steady” feature analysisKey ideaImpose a prior that higher order derivatives in the learned feature space must be small.MethodTrain a convolutional neural network with原创 2020-08-22 17:13:17 · 238 阅读 · 0 评论 -
论文笔记之6D姿态数据集:T-LESS An RGB-D Dataset for 6D Pose Estimation of Texture-less Objects
2.Related Datasets2.1 RGB-D DatasetsTexture-class objectsbenchmark: Model based training, detection and pose estimation of texture-less 3D objects in heavily cluttered scenes. In ACCV, 2012.15 texture-less objects represented by a color 3D mesh model原创 2020-08-21 10:01:54 · 889 阅读 · 0 评论 -
论文笔记之机械臂抓取:DexNet 2.0
Grasping PlanningFinding a gripper configuration that maximizes a success (or quality) metric.Method fall into wo categories based on success criteria:analytic methodsempirical (or data-driven) methodsComputer Vision Techniques in Robot GraspingAna原创 2020-08-21 10:01:10 · 695 阅读 · 0 评论 -
论文笔记之网络结构篇:RCNN
PAST: Combine multiplle low-level image features with high-level contextKey insights:CNN ---- bottom-up region proposals in order to localize and segment objectslabeled training data is scare ---- supervised pre-training for anauxiliary task, followed原创 2020-08-21 09:57:31 · 189 阅读 · 0 评论 -
超分文章记录 SRCNN-FSRCNN-ESPCN-VDCN-DRCN-RDN-LapSRN-SRDenseNet-SRGAN
Learning a Deep Convolutional Network for Image Super-Resolution(2004 ECCV )1、总结第一篇用深度学习做超分的文章,就是用深度学习来表示传统方式。结构比较简单。源码地址:SRCNN CODE2、思路先用 bicubic interpolation把图像scale到目标大小,然后通过提特征,非线性映射,然后恢复成图像。网络结构代码表示:import tensorflow as tf from tensorflow.c原创 2020-08-20 20:30:34 · 1969 阅读 · 0 评论 -
论文笔记:ShapeStacks
Related WorkCognitive scienceLearning physics from visual observationsStability predictionEnd-to-end approcach3.The ShapeStacks Dataset3.1.Dataset ContentEvery recorded image carries a binary stability label. Also, every image is aligned with a s原创 2020-08-20 09:46:42 · 213 阅读 · 0 评论 -
论文笔记:Motion feature network
ModelMotion Feature NetworkMotion filterTwo ways to aggregate spatial and temporal information from appearance block and motion filter.原创 2020-08-20 09:45:47 · 148 阅读 · 0 评论 -
论文笔记:Massively Parallel Video NetworksUntitled
1.IntroductionPipelining schemes tailored to sequence models (we call this predictive depth-parallelism)Show how such architectures can be augmented using multi-rate clocks and how they benefit from skip connections.It is possible to get better paralle原创 2020-08-20 09:44:40 · 244 阅读 · 0 评论 -
论文笔记:FlowNet Learning Optical Flow with Convolutional Networks
AbstractSolving the optical flow estimation problem as a supervised learning taskCompare two architectures: a generic architecture and another one including a layer that correlates feature vectors at different image locations.IntroductionTraining CNNs原创 2020-08-20 09:42:21 · 473 阅读 · 0 评论 -
论文笔记:Deep Learning from Temporal Coherence in Video
AbstractThe coherence is used as a supervisory signal over the unlabeled data, and is used to improve the performance on a supervised task of interest.work on some pose invariant object and face recognition tasks1、IntroductionAnalyse & ComparisonC原创 2020-08-13 09:40:58 · 384 阅读 · 0 评论 -
论文笔记:Closing the Loop for Robotic Grasping
Grasping Unknown ObjectsClosed-Loop GraspingVisual ServoingAdvantagesAdapt to dynamic environmentsNot necessarily require fully accurate camera calibration or position controlDrawbackTypically rely on hand-crafted image features for object dete原创 2020-08-13 09:36:57 · 667 阅读 · 0 评论 -
论文笔记:SSD-6D
Introductionthe accuracy of both detection and pose estimation hinges on three aspects:(1) the coverage of the 6D pose space in terms of viewpoint and scale,(2) the discriminative power of the feaures to tell objects and views apart(3) the robustness原创 2020-08-13 09:34:11 · 320 阅读 · 0 评论 -
论文笔记:g Synergies between Pushing and Grasping with Self-supervised Deep Reinforcement Learning
The key aspects ofsystem are:We learn joint pushing and grasping policies through self-supervised trial and error. Pushing actions are useful only if, in time, enable grasping. This is in contrast to prior approaches that define heuristics or hard-coded原创 2020-08-13 09:33:41 · 618 阅读 · 0 评论 -
论文笔记:7DOF
4.1 The set-up of simulation environmentIn order to meet the demand for a large dataset, we proposed a new method to generate images and ground truth labels from CAD models by simulation. Then we use the method to ShapeNetSem, a subset of ShapeNet , which原创 2020-08-13 09:32:10 · 271 阅读 · 0 评论 -
论文笔记:VGG设计理解
Very small (3 × 3) convolution filters1.IntroductionUtilised smaller receptive window size and smaller stride of the first convolutional layer.Training and testing the networks densely over the whole image and over multiple scales.Address another impo原创 2020-08-11 20:28:42 · 233 阅读 · 0 评论 -
论文笔记:ResNet
1.IntroductionProblem:Vanishing/exploding gradientsAddressed by normalized initialization [23, 9, 37, 13] and intermediate normalization layers [16], which enable networks with tens of layers to start converging for stochastic gradient descent (SGD) wi原创 2020-08-11 20:26:34 · 183 阅读 · 0 评论