论文SlimYOLOv3: Narrower, Faster and Better for Real-Time UAV Applications

分享一篇新学习的开源论文

原文：https://arxiv.org/pdf/1907.11093v1.pdf

代码：https://github.com/PengyiZhang/SlimYOLOv3

Abstract：Drones or general Unmanned Aerial Vehicles (UAVs), endowed with computer vision function by onboard cameras and embedded systems, have become popular in a wide range of applications. However,
real-time scene parsing through object detection running on a UAV platform is very challenging, due to
limited memory and computing power of embedded devices. To deal with these challenges, in this paper
we propose to learn efficient deep object detectors through channel pruning of convolutional layers. To
this end, we enforce channel-level sparsity of convolutional layers by imposing L1 regularization on
channel scaling factors and prune less informative feature channels to obtain “slim” object detectors.
Based on such approach, we present SlimYOLOv3 with fewer trainable parameters and floating point
operations (FLOPs) in comparison of original YOLOv3 (Joseph Redmon et al., 2018) as a promising
solution for real-time object detection on UAVs. We evaluate SlimYOLOv3 on VisDrone2018-Det
benchmark dataset; compelling results are achieved by SlimYOLOv3 in comparison of unpruned
counterpart, including ~90.8% decrease of FLOPs, ~92.0% decline of parameter size, running ~2 times
faster and comparable detection accuracy as YOLOv3. Experimental results with different pruning ratios
consistently verify that proposed SlimYOLOv3 with narrower structure are more efficient, faster and
better than YOLOv3, and thus are more suitable for real-time object detection on UAVs. Our codes are
made publicly available at https://github.com/PengyiZhang/SlimYOLOv3.
翻译：无人机或一般无人驾驶飞机(UAVs)通过机载摄像头和嵌入式系统赋予了计算机视觉功能，已经在广泛的应用中受到欢迎。然而,在无人机平台上通过目标检测进行实时场景解析非常具有挑战性嵌入式设备有限的内存和计算能力。为了应对这些挑战，在本文中我们建议通过卷积层的通道剪枝来学习高效的深目标检测器。来为此，我们通过施加L1正则化来增强卷积层的信道级稀疏性通道尺度因子和信息较少的特征通道的剪枝，以获得“瘦”的对象检测器。在此基础上，我们提出了具有较少可训练参数和浮点数的SlimYOLOv3与原YOLOv3 (Joseph Redmon等，2018)相比，operations (FLOPs)很有前途无人机实时目标检测解决方案。我们在VisDrone2018-Det上评估SlimYOLOv3基准数据集;与未修剪的相比，SlimYOLOv3获得了令人信服的结果对应，包括90.8%的FLOPs下降，92.0%的参数尺寸下降，运行~2次与YOLOv3相比，速度更快，检测精度也更好。不同修剪比例下的实验结果一致地验证了所提出的具有更窄结构的SlimYOLOv3是更有效、更快和更安全的优于YOLOv3，更适合无人机实时目标检测。

---

---

引用博客：https://www.jiqizhixin.com/articles/2019-07-29-10

---

---

上图，在VisDrone2018-Det基准数据集上，十亿个浮点运算(BFLOPs)与精度(mAP)的对比。通过通道剪枝（通道减少），我们的SlimYOLOv3-SPP3实现了与YOLOv3相当的检测精度，但是只需要与YOLOv3-tiny相同的浮点运算。这种性能在无人机应用中非常有竞争力。

---

---

剪枝

剪枝顾名思义，就是删去一些不重要的节点，来减小计算或搜索的复杂度。剪枝在很多算法中都有很好的应用，如：决策树，神经网络，搜索算法,数据库的设计等。在决策树和神经网络中，剪枝可以有效缓解过拟合问题并减小计算复杂度；在搜索算法中，可以减小搜索范围，提高搜索效率

---

---

一种典型的增量模型剪枝方法。存在四个迭代步骤:(1)评估一个预先训练的深度模型中每个组件的重要性;(2)剔除对模型推理不重要的成分;(3)微调修剪模型，以弥补潜在的暂时性能下降;(4)对微调后的模型进行评估，确定修剪后的模型是否适合部署。为了防止过度剪枝，最好采用增量剪枝策略。

针对SlimYOLOv3，提出了一种通过稀疏训练和通道剪枝来学习高效深目标探测器的迭代方法

YOLOv3-SPP3中SPP模块的体系结构。我们在每个检测头前面的第5和第6个卷积层之间集成了YOLOv3中的SPP模块，形成了YOLOv3- spp3。

实验结果与代码再原文中均有体现，可自行查看！

conclusion
在本文中，我们提出通过卷积层的通道剪枝来学习高效的深目标检测器。为此，我们通过在信道尺度因子上施加L1正则化来增强卷积层的信道级稀疏性，并通过对具有小尺度因子的信息较少的特征信道进行修剪来获得“瘦”的对象检测器。在此基础上，我们进一步提出了结构更窄、可训练参数更少的SlimYOLOv3。我们的SlimYOLOv3能够达到与YOLOv3相当的检测精度，而且显著减少了错误，运行速度更快。众所周知，功耗总是与FLOPs呈正相关，无人机应用一般需要低功耗才能保证无人机的续航能力。因此，我们认为SlimYOLOv3在实时UVA应用方面比原来的YOLOv3更快、更好。
In this paper, we propose to learn efficient deep object detectors through channel pruning of convolutional layers. To this end, we enforce channel-level sparsity of convolutional layers by imposing L1 regularization on the channel scaling factors and prune the less informative feature channels with small scaling factors to obtain “slim” object detectors. Based on such approach, we further present SlimYOLOv3 with narrower structure and fewer trainable parameters than YOLOv3. Our SlimYOLOv3 is able to achieve comparable detection accuracy as YOLOv3 with significantly fewer FLOPs and run faster. As known to us all, power consumption is always positively correlated with FLOPs and low power consumption is generally required by drone applications to ensure endurance of drones. Therefore, we argue that SlimYOLOv3 is faster and better than original YOLOv3 for real-time UVA applications.