【目标检测-YOLO】PP-YOLO(v3版本)

本文下载的论文是 arXiv:2007.12099v3 [cs.CV] 3 Aug 2020

PP-YOLO主要是在 YOLOv3的基础上进行的各种优化。

PP-YOLO: An Effective and Effificient Implementation of Object Detector

百度

摘要

Due to limitation of hardware, it is often necessary to sacrififice(牺牲) accuracy to ensure the infer speed of the detector in practice. Therefore, the balance between effectiveness(前文中的accuracy ) and effificiency of object detector must be considered.

We mainly try to combine various existing tricks that almost not increase the number of model parameters and FLOPs, to achieve the goal of improving the accuracy of detector as much as possible while ensuring that the speed is almost unchanged.Since all experiments in this paper are conducted based on PaddlePaddle, we call it PP-YOLO.

By combining multiple tricks, PP-YOLO can achieve a better balance between effectiveness (45.2% mAP) and effificiency (72.9 FPS), surpassing the existing state-of-the art detectors such as EffificientDet and YOLOv4. Source code is at https://github.com/PaddlePaddle/ PaddleDetection.

关键词：actual application scenarios

1. Introduction

remarkable：引人注目的；  overall：整体； nor： 也没； intended to：打算做；  recipe： 菜谱

YOLOv4 ：

• different backbone networks and data augmentation methods
• NAS to search for hyperparameters

PP-YOLO：

• backbone：ResNet
• data augmentation：MixUp
• use the manually set parameters following YOLOv3

原因：

1. ResNet 广泛使用，大量的深度学习框架对 ResNet 系列进行了深度优化，部署方便+实际中推理速度快；
2. backbone 和 data augmentation的替换 与本文讨论的无关；
3. NAS搜索超参数，需要大量计算能力，没条件用啊

所以：使用更好的 backbone + data augmentation + NAS搜索超参数 可以提高  PP-YOLO 性能。

本文的集中在怎么去堆叠一些有效tricks来得到更好的性能，这些tricks几乎不影响效率（推理速度）。许多这些 tricks 不能直接应用到 YOLOv3架构，需要做一些小的修改。此外，在哪里add tricks需要小心考虑和试验。We have found some tricks that are effective for the YOLOv3 detector, which can save developers’ time of trial and error. 降低试错时间。

2. Related Work

evolved：进化；promote：促进；reformulate：换个说法讲

Breaking the limitation imposed by hand-craft anchors, anchor-free methods show great potential for extreme object scales and aspect ratios. The performance of some recently proposed anchor-free detectors can also compete with state-of-the-art anchor-based detectors. Many tricks we discussed in this paper are different from YOLOV4 and the detailed implementation of tricks is also different.

anchor-free方法在极端目标尺度和宽高比方面显示了巨大的潜力。本文中讨论的tricks 和YOLOV4 不同，细节实现也不同。

3. Method

revise: 复习；bunch：一束

1. 替换YOLOv3的backbone 为：ResNet50-vd-dcn；
2. a bunch of tricks提高性能但不损失效率。

3.1. Architecture

extensively：广泛地；deformable：可变形；inject：添加；omitted：省略；brevity：简洁；

lateral：横向；entropy：熵； correspondingly：相应地； supervise：监督；

 

 三角形 为 DropBlock；钻石为 CoordConv；五角星为 SPP。

好吧，这个图确实很烂，看不到细节。有时间自己画下。

3.2. Selection of Tricks

accordingly：相应地；maintain：保持；Exponential：指数； perspective：透视的；translation invariance：平移不变性；integrates：集成；

 显然这些Tricks中，个人认为 Grid Sensitive、Matrix NMS、IoU Aware 可以尝试下，别的一般模型都有了。

待续......