不知不觉2020年“计算机视觉战队”陪伴大家快两个月了,由于疫情大家最近估计都没有吃好喝好,但是大家肯定玩的很High,我们也一直在陪伴,分享最好最有质量的知识,陪伴大家度过疫情。今天开始,我们准备分享一次综述性知识,有兴趣的同学加入我们一起来学习,共同进步!
目 录
1 INTRODUCTION
2 OBJECT DETECTION IN 20 YEARS
2.1 A Road Map of Object Detection
2.1.1 Milestones: Traditional Detectors
2.1.2 Milestones: CNN based Two-stage Detectors
2.1.3 Milestones: CNN based One-stage Detectors
2.2 Object Detection Datasets and Metrics
2.2.1 Metrics
2.3 Technical Evolution in Object Detection
2.3.1 Early Time’s Dark Knowledge
2.3.2 Technical Evolution of Multi-Scale Detection
2.3.3 Technical Evolution of Bounding Box Regression
2.3.4 Technical Evolution of Context Priming
2.3.5 Technical Evolution of Non-Maximum Suppression
2.3.6 Technical Evolution of Hard Negative Mining
3 SPEED-UP OF DETECTION
3.1 Feature Map Shared Computation
3.1.1 Spatial Computational Redundancy and Speed Up
3.1.2 Scale Computational Redundancy and Speed Up
3.2 Speed up of Classifiers
3.3 Cascaded Detection
3.4 Network Pruning and Quantification
3.4.1 Network Pruning
3.4.2 Network Quantification
3.4.3 Network Distillation
3.5 Lightweight Network Design
3.5.1 Factorizing Convolutions
3.5.2 Group Convolution
3.5.3 Depth-wise Separable Convolution
3.5.4 Bottle-neck Design
3.5.5 Neural Architecture Search
3.6 Numerical Acceleration
3.6.1 Speed Up with Integral Image
3.6.2 Speed Up in Frequency Domain
3.6.3 Vector Quantization
3.6.4 Reduced Rank Approximation
4 RECENT ADVANCES IN OBJECT DETECTION
4.1 Detection with Better Engines&Object detectors with new engines
4.2 Detection with Better Features
4.2.1 Why Feature Fusion is Important?
4.2.2 Feature Fusion in Different Ways
4.2.3 Learning High Resolution Features with Large Receptive Fields
4.3 Beyond Sliding Window
4.4 Improvements of Localization
4.4.1 Bounding Box Refinement
4.4.2 Improving Loss Functions for Accurate Localization
4.5 Learning with Segmentation
4.5.1 Why Segmentation Improves Detection?
4.5.2 How Segmentation Improves Detection?
4.6 Robust Detection of Rotation and Scale Changes
4.6.1 Rotation Robust Detection
4.6.2 Scale Robust Detection
4.7 Training from Scratch
4.8 Adversarial Training
4.9 Weakly Supervised Object Detection
5 APPLICATIONS
5.1 Pedestrian Detection
5.1.1 Difficulties and Challenges
5.1.2 Literature Review
5.2 Face Detection
5.2.1 Difficulties and Challenges
5.2.2 Literature review
5.3 Text Detection
5.3.1 Difficulties and Challenges
5.3.2 Literature Review
5.4 Traffic Sign and Traffic Light Detection
5.4.1 Difficulties and Challenges
5.4.2 Literature Review
5.5 Remote Sensing Target Detection
5.5.1 Difficulties and Challenges
5.5.2 Literature Review
6 CONCLUSION AND FUTURE DIRECTIONS
今天我们就先说说第一章:INTRODUCTION
目标检测是数字图像中某一类 ( 如人、动物或汽车 ) 的重要计算机视觉任务。目标检测的目标是开发计算模型和技术,提供计算机视觉应用程序所需的最基本的信息之一:什么目标在哪里?
目标检测作为计算机视觉的基本问题之一,是许多其他计算机视觉任务的基础,如实例分割、图像字幕、目标跟踪等。从应用程序的角度来看,目标检测可以被分为两个研究主题:“ General Object Detection ” 和 “ Detection Applications ” ,前者旨在探索在统一的框架下检测不同类型物体的方法,以模拟人类的视觉和认知;后者是指特定应用场景下的检测,如行人检测、人脸检测、文本检测等。
近年来,随着深度学习技术的快速发展,为目标检测注入了新的血液,取得了显著的突破,将其推向了一个前所未有的研究热点。目前,目标检测已广泛应用于自主驾驶、机器人视觉、视频监控等领域。下图就显示了过去二十年中与 “ 目标检测 ” 相关的出版物数量的增长。
区别
近年来发表了许多关于 General Object Detection 的综述。本文与上述综述的主要区别总结如下:
难点和挑战
尽管人们总是问 “ 在目标检测中有哪些困难和挑战? ” ,事实上,这个问题并不容易回答,甚至可能被过度概括。由于不同的检测任务具有完全不同的目标和约束,它们的困难程度可能会有所不同。除了其他计算机视觉任务中的一些常见挑战,如不同视点下的物体、光照和类内变化,目标检测的挑战包括但不限于以下几个方面:目标旋转和尺度变化 ( 如小目标 ) ,精确的目标定位,密集和遮挡的目标检测,加速检测等。
在之后的第四章和第五章中,我们将对这些主题进行更详细的分析。