|
|
@@ -1,172 +0,0 @@
|
|
|
-# [PaddleRS：无人机汽车识别](https://aistudio.baidu.com/aistudio/projectdetail/3713122)
|
|
|
-
|
|
|
-基于0.5m的高分辨率无人机影像，我们希望能够使用目标检测的方法找到影像中的汽车。项目将基于PaddleRS完成该任务。
|
|
|
-
|
|
|
-## 1 数据准备
|
|
|
-
|
|
|
-数据来自于[DFC2018 Houston](https://hyperspectral.ee.uh.edu/?page_id=1075)，裁剪为1400张596x601大小的图块，由手工标注而成并按照9:1划分训练集和数据集。
|
|
|
-
|
|
|
-```python
|
|
|
-# 解压数据集
|
|
|
-! mkdir -p dataset
|
|
|
-! unzip -oq data/data56250/carDetection_RGB.zip -d dataset
|
|
|
-```
|
|
|
-
|
|
|
-```python
|
|
|
-# 划分数据集
|
|
|
-import os
|
|
|
-import os.path as osp
|
|
|
-import random
|
|
|
-
|
|
|
-def 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")
|
|
|
-```
|
|
|
-
|
|
|
-## 2 PaddleRS准备
|
|
|
-
|
|
|
-PaddleRS是基于飞桨开发的遥感处理平台，支持遥感图像分类，目标检测，图像分割，以及变化检测等常用遥感任务，帮助开发者更便捷地完成从训练到部署全流程遥感深度学习应用。
|
|
|
-
|
|
|
-github：[https://github.com/PaddleCV-SIG/PaddleRS](https://github.com/PaddleCV-SIG/PaddleRS)
|
|
|
-
|
|
|
-```python
|
|
|
-! git clone https://github.com/PaddleCV-SIG/PaddleRS.git
|
|
|
-! pip install -q -r PaddleRS/requirements.txt
|
|
|
-
|
|
|
-import sys
|
|
|
-sys.path.append("PaddleRS")
|
|
|
-```
|
|
|
-
|
|
|
-## 3 模型训练
|
|
|
-
|
|
|
-PaddleRS借鉴PaddleSeg的API设计模式并进行了较高程度的封装，可以方便的完成数据、模型等的定义，快速开始模型的训练迭代。
|
|
|
-
|
|
|
-### 3.1 数据定义
|
|
|
-
|
|
|
-主要通过`datasets`和`transforms`两个组件完成任务，`datasets`中有包含分割检测分类等多任务的数据加载API，而`transforms`集成了大部分通用或单独的数据增强API，目前可以通过源码查看。
|
|
|
-
|
|
|
-```python
|
|
|
-import os
|
|
|
-import os.path as osp
|
|
|
-from paddlers.datasets import VOCDetection
|
|
|
-from paddlers import transforms as T
|
|
|
-
|
|
|
-# 定义数据增强
|
|
|
-train_transforms = T.Compose([
|
|
|
- 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])
|
|
|
-])
|
|
|
-eval_transforms = T.Compose([
|
|
|
- T.Resize(target_size=608, interp='CUBIC'),
|
|
|
- T.Normalize(
|
|
|
- mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
|
|
|
-])
|
|
|
-
|
|
|
-# 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 = VOCDetection(
|
|
|
- data_dir=data_dir,
|
|
|
- file_list=train_file_list,
|
|
|
- label_list=label_file_list,
|
|
|
- transforms=train_transforms,
|
|
|
- shuffle=True)
|
|
|
-eval_dataset = VOCDetection(
|
|
|
- data_dir=data_dir,
|
|
|
- file_list=train_file_list,
|
|
|
- label_list=label_file_list,
|
|
|
- transforms=eval_transforms,
|
|
|
- shuffle=False)
|
|
|
-```
|
|
|
-
|
|
|
-### 3.2 模型准备
|
|
|
-
|
|
|
-PaddleRS将模型分别放置于`models`和`custom_models`中，分别包含了Paddle四大套件的模型结构以及与遥感、变化检测等相关的模型结构。通过`tasks`进行了模型的封装，集成了Loss、Opt、Metrics等，可根据需要进行修改。这里以默认的PPYOLOv2为例。
|
|
|
-
|
|
|
-```python
|
|
|
-from paddlers.tasks.object_detector import PPYOLOv2
|
|
|
-
|
|
|
-num_classes = len(train_dataset.labels)
|
|
|
-model = PPYOLOv2(num_classes=num_classes)
|
|
|
-```
|
|
|
-
|
|
|
-```python
|
|
|
-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)
|
|
|
-```
|
|
|
-
|
|
|
-## 4 模型评估
|
|
|
-
|
|
|
-只需要调用evaluate即可完成预测。
|
|
|
-
|
|
|
-```python
|
|
|
-model.evaluate(eval_dataset)
|
|
|
-```
|
|
|
-
|
|
|
-返回如下输出。
|
|
|
-
|
|
|
-```
|
|
|
- 2022-03-30 19:59:13 [INFO] Start to evaluate(total_samples=944, total_steps=944)...
|
|
|
- 2022-03-30 20:00:05 [INFO] Accumulating evaluatation results...
|
|
|
-
|
|
|
- OrderedDict([('bbox_map', 90.33284968764544)])
|
|
|
-```
|
|
|
-
|
|
|
-## 5 模型预测
|
|
|
-
|
|
|
-PaddleRS的目标检测task可以方便的给出坐标、类别和分数，可供自行进行一些后处理。也可以直接使用visualize_detection进行可视化。下面对一张测试图像进行预测并可视化。
|
|
|
-
|
|
|
-```python
|
|
|
-from paddlers.tasks.utils.visualize import visualize_detection
|
|
|
-import matplotlib.pyplot as plt
|
|
|
-
|
|
|
-%matplotlib inline
|
|
|
-
|
|
|
-img_path = "dataset/JPEGImages/UH_NAD83_272056_3289689_58.jpg"
|
|
|
-pred = model.predict(img_path, eval_transforms)
|
|
|
-vis_img = visualize_detection(img_path, pred, save_dir=None)
|
|
|
-plt.figure(figsize=(10, 10))
|
|
|
-plt.imshow(vis_img)
|
|
|
-plt.show()
|
|
|
-```
|
|
|
-
|
|
|
-
|
|
|
-
|
|
|
-## 总结
|
|
|
-
|
|
|
-- 这里PPYOLOv2的效果很不错，后续在目标检测方面，将会为PaddleRS增加滑框预测以及GeoJSON等数据格式的导出。
|
geoyee