NMS by Representative Region: Towards Crowded Pedestrian Detection by
Proposal Pairing论文笔记

看论文的笔记，欢迎一起讨论，毕竟我是条咸鱼，求大佬指点*

一般学习的过程就是发现问题,解决问题，验证自己是对的。

1.发现问题：

pedestrian detection in crowded scenes is still challenging
然后作者发现目前的NMS方法不太work的样子，然后就提出了a novel Representative Region NMS (R^2NMS)的方法

目前NMS的缺点有：
A relative low threshold of intersection over union (IoU) leads to missing highly over lapped pedestrians, while a higher one brings in plenty of false positives.
这句话说IOU太低的话过度重叠部分会检测不到，过高会带来假正例的比较多.
对NMS和Adaptive NMS的分析。
*对NMS的分析

红色的框是full body predictions
绿色框是visible body predictions
原始的NMS方法会让红色虚点框消失, 而作者的方法保留了这点.

*对Adaptive NMS的分析

绿色框是真实的
红色框是预测的
预测过高,将会删除过多的预测框
However, density estima tion itself remains a diffificult task.
Besides, the matching from the density to the optimal IoU threshold is still handcrafted in AdaptiveNMS, and thus the exact matching is diffificult to acquire.

解决问题：

这里的解决问题指作者提出了怎样的模型

作者的基本框架如图, CNN Feature Extractor 在 CrowdHuman dataset中作者利用 Feature Pyramid Network (FPN) with ResNet-50 作为baseline
对于 CityPersons dataset, 采取了Faster R-CNN framework.

对于这一部分的设计，作者给出了具体算法

这个算法就是看visual part,当这一部分超过IOU留下visual part 和Full part
同时满足

其中A是 an anchor A is viewed as positive matched to the
ground-truth pair Q = (F, V)
α1 = 0.7 and β1 = 0.7.

这一部分是将特征级联,但是中间加入01色素图与Full part 做元素相乘
for a pair of annotation Q = (F, V), a pair of proposal X = (Pf,Pv) is positive

α2, β2 is 0.5 and 0.5.

验证问题：

具体的实验步骤，就是比之前的都改进了一点点
对于整篇论文，我感觉就是安照这个模型做，然后就没什么想法，
没什么改进, 完全不知道如何改进
如果对问题重定义，将Crowded Pedestrian Detection改为
Crowded Pedestrian search 是不是也有研究意义

NMS by Representative Region: Towards Crowded Pedestrian Detection by Proposal Pairing论文笔记相关推荐

(论文阅读)NMS by Representative Region: Towards Crowded Pedestrian Detection by Proposal Pairing
2020CVPR的密集行人检测论文原文链接: NMS by Representative Region: Towards Crowded Pedestrian De ...
NMS by Representative Region: Towards Crowded Pedestrian Detection by Proposal Pairing
NMS by Representative Region: Towards Crowded Pedestrian Detection by Proposal Pairing Abstract 1. I ...
Deep Anomaly Detection with Deviation Networks 论文笔记
Deep Anomaly Detection with Deviation Networks 论文笔记(自己回顾用) 问题描述算法框架伪代码总结&个人计划问题描述论文利用少部分有标签 ...
Single-Shot Object Detection with Enriched Semantics 论文笔记
前言在SSD中,小目标由较浅的层来检测,大目标由较深的层来检测.SSD虽然能够检测不同尺度的目标,但由于浅层特征只能捕获一些基础的视觉模式,语义信息不充足,因此在检测小目标时的性能不是太好:而浅层特 ...
Attention-Based Recurrent Neural Network Models for Joint Intent Detection and Slot Filling论文笔记
文章目录摘要方法 Encoder-Decoder Model with Aligned Inputs Attention-Based RNN Model 实验论文连接:Attention-Bas ...
ECCV 2020 DETR:《End-to-End Object Detection with Transformers》论文笔记
目录简介动机贡献方法实验简介本文出自FaceBook AI,方法很新颖,不同于已有的R-CNN一系列,也不同于Yolo一系列,也不同于FCOS.CenterNet等最新的Anchor F ...
Unifying Voxel-based Representation with Transformer for 3D Object Detection （UVTR）论文笔记
原文链接:https://arxiv.org/abs/2206.00630 1.引言统一表达对多模态的知识迁移和特征融合是必不可少的. 目前的统一表达可大致分为输入级和特征级的.输入级方法在网络开始 ...
F2DNet: Fast Focal Detection Network for Pedestrian Detection
F2DNet 论文标题:F2DNet: Fast Focal Detection Network for Pedestrian Detection 论文地址:https://arxiv.org/abs ...
【论文解读】V2F-Net: Explicit Decomposition of Occluded Pedestrian Detection（遮挡行人的分解检测）
论文题目:V2F-Net: Explicit Decomposition of Occluded Pedestrian Detection 论文出处:Computer Vision and Pat ...

NMS by Representative Region: Towards Crowded Pedestrian Detection by Proposal Pairing论文笔记

1.发现问题：

解决问题：

验证问题：

NMS by Representative Region: Towards Crowded Pedestrian Detection by Proposal Pairing论文笔记相关推荐

最新文章

热门文章