YOLOv8小目標(biāo)TinyPerson 行人數(shù)據(jù)集識(shí)別 行人小目標(biāo)檢測(cè) 包括數(shù)據(jù)準(zhǔn)備、模型訓(xùn)練、評(píng)估和推理。小目標(biāo)檢測(cè)Tinyperson數(shù)據(jù)集。其train71748以及test78130處理好的yolo格式(txt)以及voc格式(xml)標(biāo)簽訓(xùn)練時(shí)可自行混合后按比例劃分訓(xùn)練集和驗(yàn)證集。標(biāo)簽類(lèi)別0:earth_person1:sea_person包括數(shù)據(jù)準(zhǔn)備、模型訓(xùn)練、評(píng)估和可視化。整個(gè)代碼塊在一個(gè) artifact 中方便一次性復(fù)制。完整代碼importosimportcv2importnumpy as np from sklearn.model_selectionimporttrain_test_splitimportshutil from ultralyticsimportYOLOimportmatplotlib.pyplot as plt# Define pathsdata_pathpath_to_TinyPersonimages_train_pathos.path.join(data_path,images,train)images_test_pathos.path.join(data_path,images,test)labels_yolo_train_pathos.path.join(data_path,labels_yolo,train)labels_yolo_test_pathos.path.join(data_path,labels_yolo,test)# Create directories if they dont existos.makedirs(images_train_path,exist_okTrue)os.makedirs(images_test_path,exist_okTrue)os.makedirs(labels_yolo_train_path,exist_okTrue)os.makedirs(labels_yolo_test_path,exist_okTrue)# Combine train and test sets for splitting into train, validation, and testall_images[]all_labels[]forfilenameinos.listdir(images_train_path): image_filenamefilename label_filenameos.path.splitext(filename)[0].txtimage_pathos.path.join(images_train_path, image_filename)label_pathos.path.join(labels_yolo_train_path, label_filename)all_images.append(image_path)all_labels.append(label_path)forfilenameinos.listdir(images_test_path): image_filenamefilename label_filenameos.path.splitext(filename)[0].txtimage_pathos.path.join(images_test_path, image_filename)label_pathos.path.join(labels_yolo_test_path, label_filename)all_images.append(image_path)all_labels.append(label_path)# Split data into train, validation, and test setstrain_images, temp_images, train_labels, temp_labelstrain_test_split(all_images, all_labels,test_size0.3,random_state42)val_images, test_images, val_labels, test_labelstrain_test_split(temp_images, temp_labels,test_size0.5,random_state42)# Move files to respective foldersdef move_files(images, labels, dest_image_folder, dest_label_folder):forimage_path, label_pathinzip(images, labels): shutil.move(image_path, dest_image_folder)shutil.move(label_path, dest_label_folder)move_files(train_images, train_labels, os.path.join(data_path,images,train), os.path.join(data_path,labels_yolo,train))move_files(val_images, val_labels, os.path.join(data_path,images,val), os.path.join(data_path,labels_yolo,val))move_files(test_images, test_labels, os.path.join(data_path,images,test), os.path.join(data_path,labels_yolo,test))# Create dataset.yaml file for YOLOv8dataset_yaml_content train: ./images/train val: ./images/val test: ./images/test nc:2names:[earth_person,sea_person] with open(os.path.join(data_path,dataset.yaml),w)as f: f.write(dataset_yaml_content)# Step 3: Train YOLOv8 Model# Load a pre-trained YOLOv8 modelmodelYOLO(yolov8n.pt)# You can choose other sizes like yolov8s, yolov8m, yolov8l, yolov8x# Modify the number of classes in the final layermodel.nc2# Training commandresultsmodel.train(dataos.path.join(data_path,dataset.yaml),imgsz640,epochs50,batch16,devicecudaiftorch.cuda.is_available()elsecpu,cacheTrue)# Evaluate the modelmetricsmodel.val()# Export the trained modelmodel.export(formatonnx)# Inference using the trained model# Load the trained modelinference_modelYOLO(runs/detect/train/weights/best.pt)# Path to your best weights# Perform inference on a sample imagesample_image_pathos.path.join(data_path,images,test,sample_image.jpg)# Replace with your sample image pathresultsinference_model(sample_image_path)# Visualize resultsdef plot_results(results, image_path): imagecv2.imread(image_path)imagecv2.cvtColor(image, cv2.COLOR_BGR2RGB)forresultinresults: boxesresult.boxes.cpu().numpy()forboxinboxes: rbox.xyxy[0].astype(int)clsint(box.cls[0])confbox.conf[0]cv2.rectangle(image,(r[0], r[1]),(r[2], r[3]),(0,255,0),2)cv2.putText(image, f{result.names[cls]} {conf:.2f},(r[0], r[1]-10), cv2.FONT_HERSHEY_SIMPLEX,0.9,(0,255,0),2)plt.figure(figsize(10,10))plt.imshow(image)plt.axis(off)plt.show()plot_results(results, sample_image_path)運(yùn)行腳本在終端中運(yùn)行以下命令來(lái)執(zhí)行整個(gè)流程python main.py總結(jié)以上文檔包含了從數(shù)據(jù)加載、預(yù)處理、模型構(gòu)建到訓(xùn)練、評(píng)估和可視化的所有步驟。希望這些詳細(xì)的信息和代碼能夠幫助你順利實(shí)施和優(yōu)化你的 TinyPerson 小目標(biāo)檢測(cè)系統(tǒng)。如果你有任何進(jìn)一步的問(wèn)題或需要更多幫助請(qǐng)隨時(shí)提問(wèn)自定義說(shuō)明數(shù)據(jù)文件路徑: 修改data_path變量以指向你的 TinyPerson 數(shù)據(jù)集路徑。圖像分辨率: 根據(jù)需要調(diào)整數(shù)據(jù)增強(qiáng)中的圖像大小例如imgsz640。超參數(shù)調(diào)整: 根據(jù)需要調(diào)整訓(xùn)練參數(shù)如epochs,batch_size等。模型選擇: 你可以選擇不同的 YOLOv8 模型大小yolov8n,yolov8s,yolov8m,yolov8l,yolov8x以適應(yīng)你的需求。推理樣本路徑: 修改sample_image_path變量以指向你要進(jìn)行推理的圖片路徑。通過(guò)這些步驟你可以靈活地使用 TinyPerson 數(shù)據(jù)集進(jìn)行小目標(biāo)檢測(cè)任務(wù)。