• Stars
    star
    575
  • Rank 77,055 (Top 2 %)
  • Language
    Python
  • License
    MIT License
  • Created about 6 years ago
  • Updated almost 4 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

使用YOLO_v3_tiny和B-CNN实现街头车辆的检测和车辆属性的多标签识别 Using yolo_v3_tiny to do vehicle or car detection and attribute's multilabel classification or recognize

Vehicle-Car-detection-and-multilabel-classification 车辆检测和多标签属性识别

一个基于Pytorch精简的框架,使用YOLO_v3_tiny和B-CNN实现街头车辆的检测和车辆属性的多标签识别。
(A precise pytorch based framework for using yolo_v3_tiny to do vehicle or car detection and attribute's multilabel classification or recognize)

效果如下: Vehicle detection and recognition results are as follows:


使用方法 Usage

python Vehicle_DC -src_dir your_imgs_dir -dst_dir your_result_dir

训练好的模型文件(包括车辆检测模型和多标签分类模型) trained models on baidu drive

Tranied models-vehicle detection
Tranied models-vehicle classification
在运行Vehicle_DC脚本之前,先下载上面的模型文件或者使用自己预先训练好的模型文件,将car_540000.weights(用于检测)放在项目根目录,将epoch_39.pth(用于多标签识别)放在根目录下的checkpoints目录下,即可使用Vehicle_DC运行。
Before running Vehicle_DC, you should download provided model files provided above or use your own pretrained models. If using models provided, you need to place car_540000.weights on root directory of this project, and place epoch_39.pth on root/checkpoints/.

程序简介 brief introductions

(1). 程序包含两大模块:
The program consists of two parts: first, car detection(only provides model loading and inference code, if you need training code, you can refer to pytorch_yolo_v3); the car attributes classiyfing(provide both training and testing code, it will predict a vehicle's body color, body direction and car type)

<1>. 车辆检测模块: 只提供检测, 训练代码可以参考pytorch_yolo_v3;
<2>. 多标签识别模块:包含车辆颜色、车辆朝向、车辆类型

将这两个模块结合在一起,可以同时实现车辆的检测和识别。以此为基础,可以对室外智能交通信息,进行一定程度的结构化信息提取。
Combining these two modules together, you can do vehicle detection and multi-label recognization at the same time. Based on this info, some structured infos in outdoor traffic scenes can be extracted.

(2). 程序模块详解 modules detailed introduction

<1>. VehicleDC.py

此模块是车辆检测和车辆多标签识别接口的封装,需要指定测试源目录和结果输出目录。主类Car_DC, 函数__init__主要负责汽车检测、汽车识别两个模型的初始化。 函数detect_classify负责逐张对图像进行检测和识别:首先对输入图像进行预处理,统一输入格式,然后,输出该图像所有的车的检测框。通过函数process_predict做nms, 坐标系转换,得到所有最终的检测框。然后,程序调用函数cls_draw_bbox,在cls_draw_bbox中,逐一处理每个检测框。首先,取出原图像检测框区域检测框对应的的ROI(region of interest), 将ROI送入车辆多标签分类器。分类器调用B-CNN算法对ROI中的车辆进行多标签属性分类。参考paper link。B-CNN主要用于训练端到端的细粒度分类。本程序对论文中的网络结构做了一定的适应性修改:为了兼顾程序的推断速度和准确度,不同于论文中采用的Vgg-16,这里的B-CNN的基础网络采用Resnet-18。
This module is responsible for interface encapsulation of vehicle detection and multi-label classification. You need to specify source directory and result directory. The main class is Car_DC. The pretrained models are loaded and initiated in function init(). In function detect_classify, each input image is pre-processed to get uniformed format, then output the raw bounding boxes for further NMS calculation and coordinates tranformation. We do classification and bounding box drawing in function cls_draw_box based on bounding box ROIs. Bilinear CNN is used for fine-grained classification, and we use resnet-18 as backbone insted of vgg-16 for trade-off of accuracy and speed.

耗时统计耗时 Time consuming

车辆检测: 单张图像推断耗时,在单个GTX 1050TI GPU上约18ms。
车辆多标签识别:单张图像推断耗时,在单个GTX TITAN GPU上约7ms,在单个GTX 1050TI GPU上约10ms。
Vehicle detection: sigle image inference cost 18ms on single GTX1050TI.
Vehicle classification: single image inference cost 10ms on single GTX1050TI.

<2>. 车辆多标签数据模块(由于保密协议等原因暂时不能公开数据集) dataset.py

训练、测试数据类别按照子目录存放,子目录名即label,Color_Direction_type,如Yellow_Rear_suv。
Vehicle类重载了data.Dataset的init, getitem, len方法:
函数__init__负责初始化数据路径,数据标签,由于数据标签是多标签类型,故对输出向量分段计算交叉熵loss即可。
函数__getitem__负责迭代返回数据和标签,返回的数据需要经过标准化等预处理;函数__len__获取数据的总数量。

<3>. 车辆多标签训练、测试模块 train_vehicle_multilabel.py

此模块负责车辆多标签的训练和测试。训练过程选择交叉熵作为损失函数,需要注意的是,由于是多标签分类,故计算loss的时候需要累加各个标签的loss,其中loss = loss_color + loss_direction + 2.0 * loss_type,根据经验,将车辆类型的loss权重放到到2倍效果较好。
另一方面,训练分为两步:(1). 冻结除了Resnet-18除全连接层之外的所有层,Fine-tune训练到收敛为止;(2).打开第一步中冻结的所有层,进一步Fine-tune训练,调整所有层的权重,直至整个模型收敛为止。

More Repositories

1

MCMOT

Real time one-stage multi-class & multi-object tracking based on anchor-free detection and ReID
Python
380
star
2

RepNet-MDNet-VehicleReID

Implementing RepNet(a two-stream multitask learning network) to do vehicle Re-identification, vehicle search(or vehicle match) with PyTorch 可用于车辆细粒度识别,车辆再识别,车辆匹配,车辆检索,RepNet/MDNet的一种PyTorch实现
Python
240
star
3

FairMOTVehicle

A fork of FairMOT used to do vehicle MOT.用于跟踪车辆的多目标跟踪, 自定义数据进行单类别多目标实时跟踪
Python
181
star
4

VideoCaption

视频的文本摘要(标注),输入一段视频,通过深度学习网络和人工智能程序识别视频主要表达的意思(Input a video output a txt decribing the video)。
Python
169
star
5

FaceRecognition

Face recognition using triplet loss, implementing FaceNet with pytorch.人脸识别项目,提供一个小型数据集用作验证,使用三元组损失函数(Triplet loss)提升准确率和泛化能力,对FaceNet进行了一种实现。
Python
127
star
6

MCMOT-ByteTrack

Python
104
star
7

YOLOV4_MCMOT

Using YOLOV4 as detector for MCMOT.
Python
103
star
8

DenseBox

Implemention of Baidu's DenseBox used for multi-task learning of object detection and landmark(key-point) localization 用PyTorch实现了百度的DenseBox并针对任意宽高比矩形(不仅限于方形)的目标检测做了优化,不仅可以输出关键点的热力图(heatmap)而且可以输出每个bbox对应关键点坐标
Python
94
star
9

MOTEvaluate

Python
21
star
10

SFM_OpenCV

A SFM project implemented with Opencv
C++
18
star
11

ByteTrack-MCMOT-TensorRT

MCMOT TensorRT deployment(C/C++) based on ByteTrack.
C++
16
star
12

Depthmap-refinement-upsampling-

Implemention of paper "Spatial-Depth Super Resolution for Range Images" with python
Python
16
star
13

SFM_PMVS_3DReconstruct_python

SFM PMVS 3D sparse to dense reconstruct in python.
Python
11
star
14

BinoCameraCalibrate

Binocular camera calibration and rectification using OpenCV.
C++
11
star
15

Algorithms

Algorithms's implementions for testing 一些机器学习,统计,三维重建等算法实现和测试
Python
8
star
16

PyScripts

Python
7
star
17

ExposureFusionPy

Implementing ExposureFusion algorithm using OpenCV and Numpy in Python3.
Python
5
star
18

MonoCameraCalibrate

C++
4
star
19

ExposureFusionCpp

C++
4
star
20

TestKalman

Python
3
star
21

StereoCalibrateRectify

C++
3
star
22

MonoDepthV1

Python
3
star
23

PytorchToCaffe

Python
3
star
24

MyNanoDet

Using custom dataset for training.
Python
3
star
25

TestOpenCVCuda

Test openCV with CUDA
C++
3
star
26

MonoDepthV2

Jupyter Notebook
3
star
27

SelfSuperviseAidedBlindIQA

SelfSuperviseAidedBlindIQA
Python
2
star
28

Stereo3DReconstruct

Python
2
star
29

MyHDRUNet

Python
2
star
30

MyEnlightenGAN

Python
2
star
31

My_CRNN

2
star
32

ColmapMVSMy

C++
2
star
33

MySlamExperiments

C++
Python
2
star
34

RealChineseLiscensePlateGenerator

Python
2
star
35

CaptainEven

1
star
36

MyMVE

C++
1
star