PaddleRS：无人机汽车识别

# 解压数据集! mkdir -p dataset
! unzip -oq data/data56250/carDetection_RGB.zip -d dataset

# 划分数据集import osimport os.path as ospimport randomdef get_data_list(data_dir):
    random.seed(666)
    mode = ["train_list", "val_list"]
    dir_path = osp.join(data_dir, "JPEGImages")
    files = [f.split(".")[0] for f in os.listdir(dir_path)]
    random.shuffle(files)  # 打乱顺序
    with open(osp.join(data_dir, f"{mode[0]}.txt"), "w") as f_tr:        with open(osp.join(data_dir, f"{mode[1]}.txt"), "w") as f_va:            for i, name in enumerate(files):                if (i % 10) == 0:  # 训练集与测试集为9:1
                    f_va.write(f"JPEGImages/{name}.jpg Annotations/{name}.xml\n")                else:
                    f_tr.write(f"JPEGImages/{name}.jpg Annotations/{name}.xml\n")
    labels = ["car"]
    txt_str = "\n".join(labels)    with open((data_dir + "/" + f"label_list.txt"), "w") as f:
        f.write(txt_str)    print("Finished!")

get_data_list("dataset")

! git clone https://github.com/PaddlePaddle/PaddleRS.git
%cd PaddleRS
! git checkout release/1.0! pip install --upgrade pip
! pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
! python setup.py install
%cd ..

import osimport os.path as ospfrom paddlers.datasets import VOCDetDataset 
from paddlers import transforms as T# 定义数据增强train_transforms = T.Compose([
    T.DecodeImg(),
    T.RandomDistort(),
    T.RandomCrop(),
    T.RandomHorizontalFlip(),
    T.BatchRandomResize(
        target_sizes=[512, 544, 576, 608, 640, 672, 704],
        interp='RANDOM'),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    T.ArrangeDetector('train')
])
eval_transforms = T.Compose([
    T.DecodeImg(),
    T.Resize(target_size=608, interp='CUBIC'),
    T.Normalize(
        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    T.ArrangeDetector('eval')
])# 定义数据集data_dir = "dataset"train_file_list = osp.join(data_dir, 'train_list.txt')
val_file_list = osp.join(data_dir, 'val_list.txt')
label_file_list = osp.join(data_dir, 'label_list.txt')
train_dataset = VOCDetDataset(
    data_dir=data_dir,
    file_list=train_file_list,
    label_list=label_file_list,
    transforms=train_transforms,
    shuffle=True)
eval_dataset = VOCDetDataset(
    data_dir=data_dir,
    file_list=train_file_list,
    label_list=label_file_list,
    transforms=eval_transforms,
    shuffle=False)

# 检查及可视化数据import numpy as npimport cv2import matplotlib.pyplot as plt

%matplotlib inline# 反归一化def in_normal(img):
    mean = np.array([0.485, 0.456, 0.406])
    std = np.array([0.229, 0.224, 0.225])
    img = (255 * (img * std + mean)).astype("uint8")    return img# 绘制矩形框def visual(img, bboxs):
    for bbox in bboxs:
        x1, y1, x2, y2 = bbox
        cv2.rectangle(img, (x1, y1), (x2, y2), (255, 0, 0), 2)    return imgfor data in train_dataset:
    img = in_normal(data["image"])
    vis = visual(img, data["gt_bbox"])
    plt.figure(figsize=(10, 10))
    plt.imshow(img)
    plt.show()    break

from paddlers.tasks.object_detector import PPYOLOv2

num_classes = len(train_dataset.labels)
model = PPYOLOv2(num_classes=num_classes)

model.train(
    num_epochs=30,
    train_dataset=train_dataset,
    train_batch_size=16,
    eval_dataset=eval_dataset,
    pretrain_weights="COCO",
    learning_rate=3e-5,
    warmup_steps=10,
    warmup_start_lr=0.0,
    save_interval_epochs=5,
    lr_decay_epochs=[10, 20],
    save_dir="output",
    use_vdl=True)

from paddlers.tasks import load_model

new_model = load_model("output/best_model")

new_model.evaluate(eval_dataset)

from paddlers.tasks.utils.visualize import visualize_detectionimport matplotlib.pyplot as plt

%matplotlib inline

test_transforms = T.Compose([
    T.DecodeImg(),
    T.Resize(target_size=608, interp='CUBIC'),
    T.Normalize(
        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    T.ArrangeDetector('test')
])

img_path = "dataset/JPEGImages/UH_NAD83_272056_3289689_58.jpg"pred = new_model.predict(img_path, test_transforms)
vis_img = visualize_detection(img_path, pred, save_dir=None)
plt.figure(figsize=(10, 10))
plt.imshow(vis_img)
plt.show()