ZeroRains Blog

计算机视觉产业落地的难点与应对 出自：百度大脑——(计算机视觉产业落地难点与应对)https://www.bilibili.com/video/BV1eB4y1T772?spm_id_from=333.999.0.0 学术研究和产业应用的区别 学术研究 产业应用 任务 1. 定义明确2. 聚焦某一任务3. 单一明确的评估标准 1. 开放问题2. 系统整体设计3. 多维度，复杂评估 数据 1. 公开数据集2. 固定不变3. 数据量大 1. 自定义数据集(也可以采用公开)2. 不断迭代3. 数据量小 算力 无硬性限制(不受到设备限制) 1. 限制硬件端部署2. 服务器部署吞吐量要求3. 实时性要求 产业落地难点 数据的规模和平衡性 数据的规模： 业务数据集通常会远小于公开数据集，通常只有几百到几千张。传统的mnist，CIFAR，VOC，COCO都比业务数据集多得多，甚至ImageNet22K，YFCC100m，JFT-300M这些数据设置达到了亿级。 数据平衡性： 理想状态：每个类别的图片数量一致 现实状态：每个类别的图片数量基本不一致，有的设置 ...

You Cannot Easily Catch Me: A Low-Detectable Adversarial Patch for ObjectDetectors 论文：[You Cannot Easily Catch Me: A Low-Detectable Adversarial Patch for ObjectDetectors](You Cannot Easily Catch Me: A Low-Detectable Adversarial Patch for ObjectDetectors) 作者：Zijian Zhu, Hang Su, Chang Liu, Wenzhao Xiang, Shibao Zheng 期刊：2021年某期刊 代码：暂时没有 原文摘要 Blind spots or outright deceit can bedevil and deceive ma-chine learning models. Unidentified objects such as digital“stickers,” also known as adversarial ...

Unifying Nonlocal Blocks for Neural Networks 论文名称：Unifying Nonlocal Blocks for Neural Networks 作者：Lei Zhu, Qi She, Duo Li, Yanye Lu, Xuejing Kang, Jie Hu, Changhu Wang 期刊：ICCV2021 代码：https://github.com/zh460045050/SNL_ICCV2021 原文摘要 The nonlocal-based blocks are designed for capturinglong-range spatial-temporal dependencies in computer vi-sion tasks. Although having shown excellent performance,they still lack the mechanism to encode the rich, struc-tured information among elements in ...

习题集 第一题 解题思路：B优先级高于A，所以B先运行，但是A运行过程中不能被中断，DEV1和DEV2 可以同时工作，忽略程序切换时间。先画出两个程序运行的时间轴，所有问题都迎刃而解。 解题：时间轴线如下图所示(记号B(x)表示，B程序占用当前设备，运行时间为x) 从图中可以看出 B程序先结束 全部执行完成需要234ms 这个要计算整个程序运行在CPU上的时间总和 利用率=20+25+20+20+10+20+45+20234=180234≈76.9%利用率 =\frac{20+25+20+20+10+20+45+20}{234} = \frac{180}{234} \approx 76.9\% 利用率=23420+25+20+20+10+20+45+20​=234180​≈76.9% 计算如下 A等待CPU的时间=20+(214−199)=20+15=35msA等待CPU的时间 =20+(214-199)=20+15=35ms A等待CPU的时间=20+(214−199)=20+15=35ms 计算如下 B等待CPU的时间=(129−110)+(169−159 ...

FaPN: Feature-aligned Pyramid Network for Dense Image Prediction 论文：FaPN: Feature-aligned Pyramid Network for Dense Image Prediction 作者：Shihua Huang, Zhichao Lu, Ran Cheng, Cheng He 期刊： ICCV2021 代码：https://github.com/EMI-Group/FaPN 原文摘要 Recent advancements in deep neural networks havemade remarkable leap-forwards in dense image prediction.However, the issue of feature alignment remains as ne-glected by most existing approaches for simplicity. Directpixel addition between upsamp ...

Global Aggregation then Local Distribution in Fully Convolutional Networks 论文：Global Aggregation then Local Distribution in Fully Convolutional Networks 作者：Xiangtai Li, Li Zhang, Ansheng You, Maoke Yang, Kuiyuan Yang, Yunhai Tong 期刊：BMVC 2019 代码：https://github.com/lxtGH/GALD-DGCNet 原文摘要 It has been widely proven that modelling long-range dependencies in fully convolutionalnetworks (FCNs) via global aggregation modules is critical for complex scene under-standing tasks such as semantic se ...

人脸识别 = 人脸检测+人脸对齐+人脸比对 活体识别目的：避免恶意使用照片对人脸识别进行攻击 活体识别的位置会根据活体检测的方法，放在人脸识别的不同阶段 放在人脸检测前，如果不是活体其他都没必要 放在人脸对齐前，如果检测不到人脸怎么知道是不是活体 放在人脸比对前，没有对齐，不能提取太确定的信息(比如动作活体检测) 活体检测攻击方法 攻击方法 解决方案 打印图片 分类器 面具 动作检测(扣眼睛嘴巴的面具可能无法其作用) 视频 分类器 活体检测方法 Colour Textrue，2016 原理：活体与非活体在RGB空间里比较难区分，但在其他颜色空间里的纹理有明显差异 算法：HSV空间人脸多级LBP特征+YCbCr空间人脸LPQ特征 Micro-texture,SSD or binocuilar depth,2018 活体检测直接放到人脸检测(SSD,MTCNN等)模块里作为一个类，即人脸检测出来的bbox里有背景，真人人脸，假人脸三类的置信度.这样可以在早期就过滤掉一部分非活体。整个系统速度非常块。后续设计的SPMT feature和TFBD fea ...

Region Mutual Information Loss for Semantic Segmentation 论文名称：Region Mutual Information Loss for Semantic Segmentation 作者：Shuai Zhao, Yang Wang, Zheng Yang, Deng Cai 期刊：未查到(2019年) 代码：https://github.com/ZJULearning/RMI 原文摘要 Semantic segmentation is a fundamental problem in computer vision. It is con-sidered as a pixel-wise classification problem in practice, and most segmentationmodels use a pixel-wise loss as their optimization criterion. However, the pixel-wise loss ignores the dependencies ...

Object-Contextual Representations for Semantic Segmentation 论文名称：Object-Contextual Representations for Semantic Segmentation 作者：Yuhui Yuan, Xilin Chen, Jingdong Wang 期刊：ECCV2020 代码：https://github.com/HRNet/HRNet-Semantic-Segmentation 原文摘要 In this paper, we study the context aggregation problem in semantic segmentation. Motivated by that the label of a pixel is the category of the object that the pixel belongs to, we present a simple yet effective approach, object-con ...

Hierarchical Multi-Scale Attention for Semantic Segmentation 论文名称：Hierarchical Multi-Scale Attention for Semantic Segmentation 作者：Andrew Tao, Karan Sapra, Bryan Catanzaro 期刊：尚未查出（时间2020） 代码：https://github.com/NVIDIA/semantic-segmentation 原文摘要 Multi-scale inference is commonly used to improve the results of semantic segmentation. Multiple images scales are passed through a network and then the results are combined with averaging or max pooling. In this work, we present an attention-based appro ...