【PyTorch】基于YOLO的多目标检测项目（一）

【PyTorch】基于YOLO的多目标检测项目（一）

【PyTorch】基于YOLO的多目标检测项目（二）

目标检测是对图像中的现有目标进行定位和分类的过程。识别的对象在图像中显示有边界框。一般的目标检测方法有两种：基于区域提议的和基于回归/分类的。这里使用一种基于回归/分类的方法，称为YOLO。

目录

准备COCO数据集

创建自定义数据集

转换数据

定义数据加载器

---

准备COCO数据集

COCO是一个大规模的对象检测，分割和字幕数据集。它包含80个对象类别用于对象检测。

下载以下GitHub存储库

https://github.com/pjreddie/darknethttps://github.com/pjreddie/darknet

创建一个名为config的文件夹，将darknet/cfg/coco.data、darknet/cfg/yolov3.cfg文件复制到config文件夹中。

创建一个名为data的文件夹，从以下链接获取coco.names文件，并将其放入data文件夹，coco.names文件包含COCO数据集中80个对象类别的列表。

darknet/data/coco.names at master · pjreddie/darknet · GitHubConvolutional Neural Networks. Contribute to pjreddie/darknet development by creating an account on GitHub.https://github.com/pjreddie/darknet/blob/master/data/coco.names将darknet/scripts/get_coco_dataset.sh文件复制到data文件夹中，并复制get_coco_cocoet.sh到data文件夹。接下来，打开一个终端并执行get_coco_cocoet.sh，该脚本将把完整的COCO数据集下载到名为coco的子文件夹中。也可通过以下链接下载coco数据集。

COCO2014_数据集-飞桨AI Studio星河社区 (baidu.com)https://aistudio.baidu.com/datasetdetail/165195

在images文件夹中，有两个名为train 2014和val 2014的文件夹，分别包含82783和40504个图像。在labels文件夹中，有两个名为train 2014和val 2014的标签，分别包含82081和40137文本文件。这些文本文件包含图像中对象的边界框坐标。此外，trainvalno5k.txt文件是一个包含117264张图像的列表，这些图像将用于训练模型。此列表是train2014和val2014中图像的组合，5000个图像除外。5k.txt文件包含将用于验证的5000个图像的列表。

创建自定义数据集

完成数据集下载后，使用PyTorch的Dataset和Dataloader类创建训练和验证数据集和数据加载器。

from torch.utils.data import Dataset
from PIL import Image
import torchvision.transforms.functional as TF
import os
import numpy as np

import torch
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(torch.__version__)

#定义CocoDataset类，并展示来自训练和验证数据集的一些示例图像
class CocoDataset(Dataset):
    def __init__(self, path2listFile, transform=None, trans_params=None):
        with open(path2listFile, "r") as file:
            self.path2imgs = file.readlines()
        
        self.path2labels = [
            path.replace("images", "labels").replace(".png", ".txt").replace(".jpg", ".txt")
            for path in self.path2imgs]

        self.trans_params = trans_params
        self.transform = transform

    def __len__(self):
        return len(self.path2imgs)
    
    def __getitem__(self, index):
        path2img = self.path2imgs[index % len(self.path2imgs)].rstrip()

        img = Image.open(path2img).convert('RGB')

        path2label = self.path2labels[index % len(self.path2imgs)].rstrip()

        labels= None
        if os.path.exists(path2label):
            labels = np.loadtxt(path2label).reshape(-1, 5)
            
        if self.transform:
            img, labels = self.transform(img, labels, self.trans_params)

        return img, labels, path2img    

root_data="./data/coco"
path2trainList=os.path.join(root_data, "trainvalno5k.txt")

coco_train = CocoDataset(path2trainList)
print(len(coco_train))

 

# 从coco_train中获取图像、标签和图像路径
img, labels, path2img = coco_train[1] 
print("image size:", img.size, type(img))
print("labels shape:", labels.shape, type(labels))
print("labels \n", labels)

path2valList=os.path.join(root_data, "5k.txt")
coco_val = CocoDataset(path2valList, transform=None, trans_params=None)
print(len(coco_val))

img, labels, path2img = coco_val[7] 
print("image size:", img.size, type(img))
print("labels shape:", labels.shape, type(labels))
print("labels \n", labels)

import matplotlib.pylab as plt
import numpy as np
from PIL import Image, ImageDraw, ImageFont
from torchvision.transforms.functional import to_pil_image
import random
%matplotlib inline
path2cocoNames="./data/coco.names"
fp = open(path2cocoNames, "r")
coco_names = fp.read().split("\n")[:-1]
print("number of classese:", len(coco_names))
print(coco_names)

def rescale_bbox(bb,W,H):
    x,y,w,h=bb
    return [x*W, y*H, w*W, h*H]
COLORS = np.random.randint(0, 255, size=(80, 3),dtype="uint8")
# fnt = ImageFont.truetype('Pillow/Tests/fonts/FreeMono.ttf', 16)
fnt = ImageFont.truetype('arial.ttf', 16)
def show_img_bbox(img,targets):
    if torch.is_tensor(img):
        img=to_pil_image(img)
    if torch.is_tensor(targets):
        targets=targets.numpy()[:,1:]
        
    W, H=img.size
    draw = ImageDraw.Draw(img)
    
    for tg in targets:
        id_=int(tg[0])
        bbox=tg[1:]
        bbox=rescale_bbox(bbox,W,H)
        xc,yc,w,h=bbox
        
        color = [int(c) for c in COLORS[id_]]
        name=coco_names[id_]
        
        draw.rectangle(((xc-w/2, yc-h/2), (xc+w/2, yc+h/2)),outline=tuple(color),width=3)
        draw.text((xc-w/2,yc-h/2),name, font=fnt, fill=(255,255,255,0))
    plt.imshow(np.array(img))        
np.random.seed(1)
rnd_ind=np.random.randint(len(coco_train))
img, labels, path2img = coco_train[rnd_ind] 
print(img.size, labels.shape)

plt.rcParams['figure.figsize'] = (20, 10)
show_img_bbox(img,labels)

np.random.seed(1)
rnd_ind=np.random.randint(len(coco_val))
img, labels, path2img = coco_val[rnd_ind] 
print(img.size, labels.shape)

plt.rcParams['figure.figsize'] = (20, 10)
show_img_bbox(img,labels)

转换数据

定义一个转换函数和传递给CocoDataset类的参数

def pad_to_square(img, boxes, pad_value=0, normalized_labels=True):
    w, h = img.size
    w_factor, h_factor = (w,h) if normalized_labels else (1, 1)
    
    dim_diff = np.abs(h - w)
    pad1= dim_diff // 2
    pad2= dim_diff - pad1
    
    if h<=w:
        left, top, right, bottom= 0, pad1, 0, pad2
    else:
        left, top, right, bottom= pad1, 0, pad2, 0
    padding= (left, top, right, bottom)

    img_padded = TF.pad(img, padding=padding, fill=pad_value)
    w_padded, h_padded = img_padded.size
            
    x1 = w_factor * (boxes[:, 1] - boxes[:, 3] / 2)
    y1 = h_factor * (boxes[:, 2] - boxes[:, 4] / 2)
    x2 = w_factor * (boxes[:, 1] + boxes[:, 3] / 2)
    y2 = h_factor * (boxes[:, 2] + boxes[:, 4] / 2)    
    
    x1 += padding[0] # 左
    y1 += padding[1] # 上
    x2 += padding[2] # 右
    y2 += padding[3] # 下
            
    boxes[:, 1] = ((x1 + x2) / 2) / w_padded
    boxes[:, 2] = ((y1 + y2) / 2) / h_padded
    boxes[:, 3] *= w_factor / w_padded
    boxes[:, 4] *= h_factor / h_padded

    return img_padded, boxes    

def hflip(image, labels):
    image = TF.hflip(image)
    labels[:, 1] = 1.0 - labels[:, 1]
    return image, labels

def transformer(image, labels, params):
    if params["pad2square"] is True:
        image,labels= pad_to_square(image, labels)
    
    image = TF.resize(image,params["target_size"])

    if random.random() < params["p_hflip"]:
        image,labels=hflip(image,labels)

    image=TF.to_tensor(image)
    targets = torch.zeros((len(labels), 6))
    targets[:, 1:] = torch.from_numpy(labels)
    
    return image, targets

trans_params_train={
    "target_size" : (416, 416),
    "pad2square": True,
    "p_hflip" : 1.0,
    "normalized_labels": True,
}
coco_train=CocoDataset(path2trainList,transform=transformer,trans_params=trans_params_train)

np.random.seed(100)
rnd_ind=np.random.randint(len(coco_train))
img, targets, path2img = coco_train[rnd_ind] 
print("image shape:", img.shape)
print("labels shape:", targets.shape) 

plt.rcParams['figure.figsize'] = (20, 10)
COLORS = np.random.randint(0, 255, size=(80, 3),dtype="uint8")
show_img_bbox(img,targets)

通过传递 transformer 函数来定义 CocoDataset 的一个对象来验证数据 

trans_params_val={
    "target_size" : (416, 416),
    "pad2square": True,
    "p_hflip" : 0.0,
    "normalized_labels": True,
}
coco_val= CocoDataset(path2valList,
                      transform=transformer,
                      trans_params=trans_params_val)

np.random.seed(55)
rnd_ind=np.random.randint(len(coco_val))
img, targets, path2img = coco_val[rnd_ind] 
print("image shape:", img.shape)
print("labels shape:", targets.shape) 

plt.rcParams['figure.figsize'] = (20, 10)
COLORS = np.random.randint(0, 255, size=(80, 3),dtype="uint8")
show_img_bbox(img,targets)

 

定义数据加载器

定义两个用于训练和验证数据集的数据加载器，从coco_train和coco_val中获取小批量数据。

from torch.utils.data import DataLoader

batch_size=8
def collate_fn(batch):
    imgs, targets, paths = list(zip(*batch))
    
    targets = [boxes for boxes in targets if boxes is not None]
    
    for b_i, boxes in enumerate(targets):
        boxes[:, 0] = b_i
    targets = torch.cat(targets, 0)
    imgs = torch.stack([img for img in imgs])
    return imgs, targets, paths

train_dl = DataLoader(
        coco_train,
        batch_size=batch_size,
        shuffle=True,
        num_workers=0,
        pin_memory=True,
        collate_fn=collate_fn,
        )

torch.manual_seed(0)
for imgs_batch,tg_batch,path_batch in train_dl:
    break
print(imgs_batch.shape)
print(tg_batch.shape,tg_batch.dtype)

 

val_dl = DataLoader(
        coco_val,
        batch_size=batch_size,
        shuffle=False,
        num_workers=0,
        pin_memory=True,
        collate_fn=collate_fn,
        )

torch.manual_seed(0)
for imgs_batch,tg_batch,path_batch in val_dl:
    break
print(imgs_batch.shape)
print(tg_batch.shape,tg_batch.dtype)