<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[风行]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/A_A666</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; A_A666]]></copyright><item><title><![CDATA[Single-Domain Generalized Object Detection in Urban Scene via Cyclic-Disentangled Self-Distill阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/125698321</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/125698321</guid><author>A_A666</author><pubDate>Sun, 10 Jul 2022 17:49:25 +0800</pubDate><description><![CDATA[Cyclic-Disentangled Self-Distillation]]></description><category></category></item><item><title><![CDATA[PKD: General Distillation Framework for Object Detectors via Pearson Correlation Coefﬁcient阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/125689284</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/125689284</guid><author>A_A666</author><pubDate>Sat, 09 Jul 2022 12:08:50 +0800</pubDate><description><![CDATA[PKD论文笔记]]></description><category></category></item><item><title><![CDATA[DN-DETR: Accelerate DETR Training by Introducing Query DeNoising阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/124007421</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/124007421</guid><author>A_A666</author><pubDate>Fri, 15 Apr 2022 10:54:58 +0800</pubDate><description><![CDATA[DN-DETR阅读笔记]]></description><category></category></item><item><title><![CDATA[DAB-DETR:Dynamic Anchor Boxes Are Better Queries for DETR阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/123483204</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/123483204</guid><author>A_A666</author><pubDate>Fri, 18 Mar 2022 19:21:37 +0800</pubDate><description><![CDATA[DAB-DETR阅读笔记]]></description><category></category></item><item><title><![CDATA[A ConvNet for the 2020s论文阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/122559336</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/122559336</guid><author>A_A666</author><pubDate>Wed, 19 Jan 2022 15:45:34 +0800</pubDate><description><![CDATA[ConvNeXt阅读笔记(一) Title(二) Summary(三) Research Object(四) Problem Statement(五) Method(六) Experiments6.1 ConvNeXt结构设计6.2 ImageNet实验6.3 目标检测任务和分割任务中的表现(七) Conclusions(八) Questions
(一) Title

论文地址:https://arxiv.org/abs/2201.03545
代码地址:https://github.com/facebook]]></description><category></category></item><item><title><![CDATA[旋转框resize中存在的问题]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/122498049</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/122498049</guid><author>A_A666</author><pubDate>Fri, 14 Jan 2022 17:05:11 +0800</pubDate><description><![CDATA[旋转框采用的表示方式为:
（a）长边表示法
longside format：   x_c      y_c       longside      shortside       Θ
参数介绍：

x_c与y_c：旋转矩形框的中心坐标；
longside: 顾名思义，最长的边；
shortside：与longside对应的另一条边；
Θ: 最长边到x轴逆时针旋转的夹角，逆时针方向角度为负 ，Θ∈[-180, 0)。

经过然后对图片进行resize之后,边界框的位置发生了变换,不能够同原始图像对应上，下面.]]></description><category></category></item><item><title><![CDATA[AutoSlim: Towards One-Shot Architecture Search for Channel Numbers阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/122200425</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/122200425</guid><author>A_A666</author><pubDate>Wed, 29 Dec 2021 15:23:50 +0800</pubDate><description><![CDATA[AutoSlim阅读笔记(一) Title(二) Summary(三) Research Object(四) Problem Statement(五) Method5.1 搜索通道数整体方案5.2 Slimmable Networks的训练5.3 Greedy Slimming(六) Experiments(七) Conclusion(八) Notes8.1 得到最优的Channel Numbers的方法
(一) Title

论文地址:https://arxiv.org/abs/1903.11728
代码]]></description><category></category></item><item><title><![CDATA[Channel-wise Knowledge Distillation for Dense Prediction阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/122137052</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/122137052</guid><author>A_A666</author><pubDate>Tue, 28 Dec 2021 17:16:03 +0800</pubDate><description><![CDATA[Channel-wise KD阅读笔记(一) Title(二) Summary(三) Research Object(四) Problem Statement(五) Method5.1 spatial distillation5.2 Channel-wise Distillation(六) Experiments6.1 实验设置6.2 和最近knowledge Distillation方法相比6.3 消融实验6.4 目标检测任务(七) Conclusion(八) Notes8.1 Dense Predict]]></description><category></category></item><item><title><![CDATA[Gaussian Bounding Boxes and Probabilistic Intersection-over-Union for Object Detection阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/121833048</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/121833048</guid><author>A_A666</author><pubDate>Tue, 21 Dec 2021 15:57:53 +0800</pubDate><description><![CDATA[ProbIoU论文阅读笔记]]></description><category></category></item><item><title><![CDATA[Deformable DETR:Deformable Transformers for End-to-end object detection笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/121721130</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/121721130</guid><author>A_A666</author><pubDate>Sat, 04 Dec 2021 20:25:34 +0800</pubDate><description><![CDATA[(一) 本文需要解决的问题
DETR具有slow convergence以及limited feature spatial resolution
原始的DETR模型存在的问题：

训练时间长,需要500个epoch才能够收敛
对小目标检测能力相对较弱
在初始化时,Attention modules接近是均匀分布，需要长时间的训练才能将其转变为sparse meaningful locations
attention weight这边的权重计算是二次复杂度的，对于高分辨率图来说是高计算和高内存消耗的。

De]]></description><category></category></item><item><title><![CDATA[TransTrack: Multiple Object Tracking with Transformer论文阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/121570451</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/121570451</guid><author>A_A666</author><pubDate>Tue, 30 Nov 2021 16:02:47 +0800</pubDate><description><![CDATA[(一) Title

论文地址：https://arxiv.org/abs/2012.15460
代码地址：https://github.com/PeizeSun/TransTrack
前言: 本文建立了一种novel joint-detection-and-tracking模型,在一个框架中完成目标检测和跟踪。这是首篇将Transformer引入到MOT任务中
(二) Summary
研究背景

目前MOT中主要有两种方式:tracking-by-detection以及joint-detection-an]]></description><category></category></item><item><title><![CDATA[FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/121564556</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/121564556</guid><author>A_A666</author><pubDate>Fri, 26 Nov 2021 21:26:57 +0800</pubDate><description><![CDATA[(一) Title

论文地址:https://arxiv.org/abs/2111.10780
项目地址:https://github.com/lzh420202/FCOSR
(二) Summary
研究背景:

现有的基于anchor-based方法的旋转框检测算法需要手动设置预设框,引入了额外的超参数和计算。基于anchor-free的方法通常具有复杂的体系结构,不容易部署。

本文方法:

本文提出了一个一阶段的带有角度旋转的检测框架FCOSR,一看就是基于FCOS架构,并且将二维高斯分布也引入进来]]></description><category></category></item><item><title><![CDATA[Oriented Object Detection with Transformer阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/121408283</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/121408283</guid><author>A_A666</author><pubDate>Fri, 19 Nov 2021 17:24:17 +0800</pubDate><description><![CDATA[(一) Title

前言: 本文没有开源,但是最近我有实现一个基于Transformer的Detector的想法,因此还是决定看下这个论文,然后再基于DETR构建一个基于Transformer的旋转框目标检测方法,怎么说呢？公式里的每一个符号我都能看懂,但是合到一起就看不懂了,而且看完这个论文的感受很不好，有太多疑问了，希望之后能够开源代码吧。
论文地址:https://arxiv.org/abs/2106.03146
代码地址:没开源
(二) Summary
背景

DETR相比于传统的detector]]></description><category></category></item><item><title><![CDATA[MOTR: End-to-End Multiple-Object Tracking with TRansformer阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/121034322</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/121034322</guid><author>A_A666</author><pubDate>Wed, 17 Nov 2021 20:00:27 +0800</pubDate><description><![CDATA[(一) Title

论文地址:https://arxiv.org/abs/2105.03247
代码地址:https://github.com/megvii-model/MOTR
前言：最近有幸看了几篇MOT的工作,感觉目前MOT的工作有一些集中将循环神经网络引入来构建对轨迹的预测，从而实现对遮挡情况的处理,提升性能。然而我自己没有仔细研读过Transformer相关的论文,希望以此为契机,开启循环神经网络在视觉领域的研究。
(二) Summary
当前存在的问题：

多目标跟踪任务之前的做法中,往往是基]]></description><category></category></item><item><title><![CDATA[Recurrence along Depth: Deep Convolutional Neural Networks with Recurrent Layer Aggregation阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/121162388</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/121162388</guid><author>A_A666</author><pubDate>Fri, 12 Nov 2021 22:27:57 +0800</pubDate><description><![CDATA[(一) Title

论文地址：https://arxiv.org/abs/2110.11852
(二) Summary
研究背景
本文内容

本文进一步探讨了layer aggregation的问题,使得previous layer的信息能够更好地服务于current layer的特征提取。
通过引入recurrent layer aggregation(RLA),充分利用CNNs layers的顺序结构,帮助学习图像中的structural information.

(三) Research Obje]]></description><category></category></item><item><title><![CDATA[Learning to Track with Object Permanence阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/120946756</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/120946756</guid><author>A_A666</author><pubDate>Tue, 26 Oct 2021 21:37:49 +0800</pubDate><description><![CDATA[Learning to Track with Object Permanence阅读笔记(一) Tilte(二) Summary(三) Research Object(四) Problem Statement(五) Method5.1 CenterTrack框架5.2 将pairs frame扩展到video-level model5.3 怎么在完全遮挡情况下进行跟踪呢？5.4 怎么处理真实数据和生成数据之间的gap(六) Experiments6.1 数据集+评价指标6.2 实验设置6.3 消融实验6.3]]></description><category></category></item><item><title><![CDATA[ByteTrack: Multi-Object Tracking by Associating Every Detection Box阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/120841975</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/120841975</guid><author>A_A666</author><pubDate>Sat, 23 Oct 2021 21:49:37 +0800</pubDate><description><![CDATA[(一) Title

论文地址：https://arxiv.org/pdf/2110.06864.pdf
代码地址：https://github.com/ifzhang/ByteTrack
前言：截止20211019在MOT17上为SOTA,刚刚看源码的时候被项目的详细程度惊艳到了,而且猛然发现本作一作也是原来FairMOT的一作,不愧是大佬啊,忒强了
(二) Summary
背景介绍

MOT的任务是estimating bounding boxes and identities of objects。从]]></description><category></category></item><item><title><![CDATA[Alpha-Refine: Boosting Tracking Performance by Precise Bounding Box Estimation阅读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/120528222</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/120528222</guid><author>A_A666</author><pubDate>Wed, 29 Sep 2021 14:28:14 +0800</pubDate><description><![CDATA[(一)Title

前言：提高tracker的性能，即插即用，无需训练，数据结果表现强劲，????啊。
(二) Summary
研究背景：

目前trackers通过采用多阶段策略改善bounding box估计质量，不过目前方法存在的问题是：

limited precision
the coupling of different stages severely restricts the method‘s transferability


本文提出方法

提出Alpha-Refine模块，用于提升边界]]></description><category></category></item><item><title><![CDATA[GMOT-40: A Benchmark for Generic Multiple Object Tracking速读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/118399853</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/118399853</guid><author>A_A666</author><pubDate>Thu, 01 Jul 2021 23:44:16 +0800</pubDate><description><![CDATA[(一)Title

前言: 这篇文章中提到GMOT和MOT并不是相同的内容,因此,这篇文章并不是关注的重点。
代码: https://github.com/Spritea/GMOT40
(二)Keypoints
问题:

现有的研究需要跟踪目标(比如行人)的先验知识,不能很好地推广到未见过的类别上(这个问题让我感到很迷惑,不知道下面introduction有没有介绍)。而General Multiple Object Tracking(GMOT)需要很少的关于目标的先验信息(这里的第二个疑问就是GMOT和M]]></description><category></category></item><item><title><![CDATA[Alpha-Refine: Boosting Tracking Performance by Precise Bounding Box Estimation速读笔记]]></title><link>https://blog.youkuaiyun.com/A_A666/article/details/118220014</link><guid>https://blog.youkuaiyun.com/A_A666/article/details/118220014</guid><author>A_A666</author><pubDate>Thu, 01 Jul 2021 22:41:39 +0800</pubDate><description><![CDATA[(一)Title

前言: 准备开启跟踪的工作了，再不做就毕不了业了，先做文献阅读啦拉啦，欧耶！就是干！
论文地址:https://arxiv.org/abs/2007.02024
代码地址: https://github.com/MasterBin-IIAU/AlphaRefine
关键点：multiple base trackers
(二)  Summary

研究现状及问题：

最近的trackers采用multiple-stage strategy来改善边界框estimation，这些方法首先粗略地]]></description><category></category></item></channel></rss>