写在前面
中介费可真贵emmmm,不过毕竟是出国,理解万岁
和中介小姐姐聊了半个下午,对出国的方法与方式有了些理解和认识,这篇文章就总结一下出国计划和安排吧。
(总不是gap一年再出国的?)
对于图像或视频中的物体进行检测是计算机视觉中基础且重要的任务,任务的难点在于在相对复杂的场景中以实时的速度对物体进行定位与分类。在目前的物体检测算法中, YOLOv3以One-Stage的方式达到了实时检测的效果,但在精度上有所欠缺。
本文针对以上问题提出一种针对YOLOv3算法的改进方案,实现对于复杂环境下不同物体的实时检测。文章从以下三部分进行阐述。
实验表明,针对BDD100K数据集,在速度在达到实时检测级别(超过30 fps)的前提下,模型能达到58.59%的mAP(平均精度),高于同等条件下原生YOLOv3模型35.6%的mAP。本研究在保证实时检测的前提下提高了YOLOv3算法的精度,实现了改进。
关键词:YOLOv3; 特征融合; 物体检测; 实时检测
Objects detection in images or videos is a fundamental and important task in computer vision. The difficulty of the task is to locate and classify objects in real-time speed in complex scenes. In current algorithms, YOLOv3 achieves the real-time detection effect by One-Stage, but it lacks precision.
In this paper, an improved scheme for YOLOv3 algorithm is proposed for the above problems, realizing real-time detection for different objects in complex environments. This article describes the improvement from the following three parts.
Analyzed and implemented YOLOv3. This research studied the backbone feature extraction network and feature interaction network of YOLOv3, and implemented YOLOv3 through model, train, predict and detect module based on PyTorch.
Analyzed the improvement ideas. Starting from the two aspects of model and dataset. On the one hand, the improved model of the prior frame of the model YOLO layer and the network structure of the feature pyramid is proposed. On the other hand, the improvement of the class, size, resolution and picture parameters of the dataset object feature is proposed.
From the two aspects of model and dataset, under the calculation standard of MS COCO mAP-30, using the test results before improvement, the accuracy of the improved algorithm is evaluated and evaluated from the aspects of overall improvement results and partial improvement results. Improve the effect.
Experiment shows that, a model is trained that achieves 58.59%mAP(mean Average Precision) at the real-time level for detection (over 30 fps) on BDD100K dataset, higher than the native YOLOv3 model's result of 35.6% mAP. This research improves the accuracy of YOLOv3 under the premise of real-time detection.
Key words: YOLOv3; feature fusion; object detection; real-time detection
在帮女票配置opencv的环境的时候,免不了要安装opencv啦,可是安装好之后一直在报错,看上去是有很多依赖找不到。就连opencv官方提供的example也跑不起来。
最有意思的是,这些依赖明明都存在,可就偏偏找不到,踩了无数个坑。从换opencv的版本、自己编译opencv,到查这些依赖的位置、CMakeLists研究——最后找到了解决方案,原来真的是有依赖没装……
于是便想顺手总结一下这个问题——关于linux软件的安装。
不久前读了一个博主关于Emacs以及GNU的看法。
其在我为什么鄙视并抵制Emacs编辑器?中表达了对当前Emacs过高吹捧的不满1,并对GNU运动及其精神领袖做出了一些点评。
老实说,有些观点不太能苟同,我对GNU的了解不算很深刻吧,但是大致知道一些开源运动,也听鸟叔常常提起,这里就写下我自己的认识。