点击下方卡片，关注“OpenCV与AI深度学习”
视觉/图像重磅干货，第一时间送达!
https://docs.ultralytics.com/models/yolo11/?source=post_page-----a793d51cc4ba--------------------------------数据准备
https://github.com/varunpn7405/Vehicle_Accident_detection避免误导模型：对象检测模型经过训练，可以预测图像中对象的存在和位置。如果包含没有注释的图像（即没有任何标记对象的图像），则模型可能会错误地了解到许多图像不包含对象，从而降低其有效检测对象的能力。 提高训练效率：在没有注释的图像上进行训练会浪费计算资源，因为模型在处理图像时没有学习任何关于对象位置的有用信息。删除这些图像有助于将训练重点放在相关的信息数据上。 减少偏差：包含大量空白图像可能会使模型偏向于预测图像通常不包含对象，从而导致更高的假阴性率（即无法检测到存在的对象）。 防止过度拟合：对过多的空图像进行训练可能会导致模型对预测未来图像的“无对象”过于自信，这可能会损害其对存在对象的真实场景的泛化。 确保正确的损失计算：目标检测模型通常使用依赖于目标存在的损失（如分类和定位损失）。空图像可能会影响这些损失的计算方式，尤其是当模型期望每个图像至少有一个对象时，这可能会导致训练期间的不稳定。
import os,shutil# Function to check if a file is emptydef is_empty_file(file_path):return os.path.exists(file_path) and os.stat(file_path).st_size == 0image_extensions = ['.jpg', '.jpeg', '.png']path = os.getcwd()inputPar = os.path.join(path, r'dataset')outputPar = os.path.join(path, r'filtered')if not os.path.exists(outputPar):os.makedirs(outputPar)folders = os.listdir(inputPar)for folder in folders:if folder in ["test","train","valid"]:inputChild = os.path.join(inputPar, folder,"labels")outputChild1 = os.path.join(outputPar, folder,"labels")if not os.path.exists(outputChild1):os.makedirs(outputChild1)outputChild2 = os.path.join(outputPar, folder,"images")if not os.path.exists(outputChild2):os.makedirs(outputChild2)files = os.listdir(inputChild)for file in files:annotation_path = os.path.join(inputChild, file)# Check if the annotation file is emptyif not is_empty_file(annotation_path):shutil.copy(annotation_path,os.path.join(outputChild1,file))# Try to find and remove the corresponding image fileimage_name = os.path.splitext(file)[0]for ext in image_extensions:image_path = os.path.join(inputPar,folder,"images", image_name + ext)if os.path.exists(image_path):shutil.copy(image_path,os.path.join(outputChild2,image_name + ext))
import os,shutilimage_extensions = ['.jpg', '.jpeg', '.png']path = os.getcwd()inputPar = os.path.join(path, r'accident detection.v10i.yolov11')outputPar = os.path.join(path, r'accident detection.v10i.yolov11(Filtered))')if not os.path.exists(outputPar):os.makedirs(outputPar)folders = os.listdir(inputPar)clsfile = os.path.join(path, 'classes.txt')with open(clsfile) as tf:clsnames = [cl.strip() for cl in tf.readlines()]for folder in folders:if folder in ["test","train","valid"]:inputChild = os.path.join(inputPar, folder,"labels")outputChild1 = os.path.join(outputPar, folder,"labels")if not os.path.exists(outputChild1):os.makedirs(outputChild1)outputChild2 = os.path.join(outputPar, folder,"images")if not os.path.exists(outputChild2):os.makedirs(outputChild2)files = os.listdir(inputChild)for file in files:fileName, ext = os.path.splitext(file)finput = os.path.join(inputChild, file)with open(finput) as tf:Yolodata = tf.readlines()# for obj in Yolodata:if not all(int(obj.split(' ')[0])==2 or int(obj.split(' ')[0])==1 for obj in Yolodata):print(file)new_yolo_data=[]for obj in Yolodata:if not (int(obj.split(' ')[0])==2 or int(obj.split(' ')[0])==1) :new_yolo_data.append(obj)with open(os.path.join(outputChild1,file),"w") as tf:tf.writelines(new_yolo_data)image_name = os.path.splitext(file)[0]for ext in image_extensions:image_path = os.path.join(inputPar,folder,"images", image_name + ext)if os.path.exists(image_path):shutil.copy(image_path,os.path.join(outputChild2,image_name + ext))break
import osfrom PIL import Image,ImageDraw,ImageFontfont = ImageFont.truetype("arial.ttf", 15)path = os.getcwd()inputPar = os.path.join(path, r'Dataset')outputPar = os.path.join(path, r'Visualisation')if not os.path.exists(outputPar):os.makedirs(outputPar)folders = os.listdir(inputPar)cls_clr = {"Accident":"#eb0523"}clsfile = os.path.join(path, 'classes.txt')with open(clsfile) as tf:clsnames = [cl.strip() for cl in tf.readlines()]for folder in folders:if folder in ["test","train","valid"]:inputChild = os.path.join(inputPar, folder,"labels")outputChild = os.path.join(outputPar, folder)if not os.path.exists(outputChild):os.makedirs(outputChild)files = os.listdir(inputChild)for file in files:fileName, ext = os.path.splitext(file)finput = os.path.join(inputChild, file)with open(finput) as tf:Yolodata = tf.readlines()imgpath1 = os.path.join(inputPar,folder,"images", fileName + '.jpg')# imgpath2 = os.path.join(inputPar,folder,"images", fileName + '.png')if os.path.exists(imgpath1):imgpath=imgpath1# elif os.path.exists(imgpath2):# imgpath=imgpath2if os.path.exists(imgpath):print("plotting >>",fileName + '.jpg')img = Image.open(imgpath)draw = ImageDraw.Draw(img)width, height = img.sizefor obj in Yolodata:clsName = clsnames[int(obj.split(' ')[0])]xnew = float(obj.split(' ')[1])ynew = float(obj.split(' ')[2])wnew = float(obj.split(' ')[3])hnew = float(obj.split(' ')[4])label=f"{clsName}"# box sizedw = 1 / widthdh = 1 / height# coordinatesxmax = int(((2 * xnew) + wnew) / (2 * dw))xmin = int(((2 * xnew) - wnew) / (2 * dw))ymax = int(((2 * ynew) + hnew) / (2 * dh))ymin = int(((2 * ynew) - hnew) / (2 * dh))clr = cls_clr[clsName]tw, th = font.getbbox(label)[2:]# draw bbox and classname::draw.rectangle([(xmin,ymin),(xmax,ymax)],outline =clr,width=2)txtbox = [(xmin,ymin-th),(xmin+tw,ymin)]draw.rectangle(txtbox, fill=clr)draw.text((xmin,ymin-th),label,fill='white',font=font)fout = os.path.join(outputChild, imgpath.split("\\")[-1])img.save(fout)else:print(f'{imgpath} >> img not found:'
相同的 Project Config File 为
[{"name": "Accident","id": 3505802,"color": "#f80b2b","type": "any","attributes": []}]
import osimport shutilfrom concurrent.futures import ThreadPoolExecutorimport timestartTime=time.time()def copy_files(src_dir, dst_dir):"""Copy files from src_dir to dst_dir."""os.makedirs(dst_dir, exist_ok=True)for file in os.listdir(src_dir):shutil.copy(os.path.join(src_dir, file), os.path.join(dst_dir, file))def process_folder(inputPar, outPar, folder):"""Process 'images' and 'labels' subfolders within the given folder."""if folder in ["train", "valid", "test"]:inputChild = os.path.join(inputPar, folder)for subfldr in ["images", "labels"]:inputSubChild = os.path.join(inputChild, subfldr)outChild = os.path.join(outPar, subfldr, folder)if os.path.exists(inputSubChild):copy_files(inputSubChild, outChild)def main():cPath = os.getcwd()inputPar = r"data_set"outPar = os.path.join(cPath, "Dataset")folders = ["train", "valid", "test"]with ThreadPoolExecutor() as executor:executor.map(lambda folder: process_folder(inputPar, outPar, folder), folders)if __name__ == "__main__":main()endTime=time.time()print("process Completed in !!",endTime-startTime)
import yaml# Define the data configurationdata_config = {'train': '/content/data_v2/images/train', # Replace with your train directory path'val': '/content/data_v2/images/valid', # Replace with your train directory path'nc': 1, # number of classes'names': ['Accident']}# Write the configuration to a YAML filewith open('data.yaml', 'w') as file:yaml.dump(data_config, file)
pip install ultralyticsfrom ultralytics import YOLO# Load a modelmodel = YOLO("yolo11n.pt")# Train the modeltrain_results = model.train(data="data.yaml", # path to dataset YAMLepochs=100, # number of training epochs do adjust as you needimgsz=640, # training image size)
model.export(format="onnx") # creates 'best.onnx'import cv2from ultralytics import YOLOimage_path=r"Accident detection model.v2i.yolov11(Empty Filtered)\train\images\Accidents-online-video-cutter_com-_mp4-41_jpg.rf.549dce3991b2ae74ae65274cc32d8eff.jpg" # Replace with your test image pathonnx_model = YOLO("best.onnx")class_names=onnx_model.namesimage = cv2.imread(image_path)# Run inferenceresults = onnx_model(image)# Extract predictionsfor result in results:boxes = result.boxes # get bounding boxesfor box in boxes:x1, y1, x2, y2 = map(int, box.xyxy[0].tolist()) # Bounding box coordinatesconf = box.conf.item() # Confidence scoreclass_id = int(box.cls.item()) # Class ID# Prepare the label textlabel = f"{class_names[class_id]}: {conf:.2f}"# Draw the bounding box (blue color, thickness of 2)cv2.rectangle(image, (x1, y1), (x2, y2), (255, 0, 0), 2)# Draw the label above the bounding boxfont = cv2.FONT_HERSHEY_SIMPLEXlabel_size, _ = cv2.getTextSize(label, font, 0.5, 1)label_ymin = max(y1, label_size[1] + 10)cv2.rectangle(image, (x1, label_ymin - label_size[1] - 10),(x1 + label_size[0], label_ymin + 4), (255, 0, 0), -1) # Draw label backgroundcv2.putText(image, label, (x1, label_ymin), font, 0.5, (255, 255, 255), 1) # Put label text# Save the imageoutput_path = "output_image.jpg"cv2.imwrite(output_path, image)print(f"Saved inference result to {output_path}")
{"image_name_1": {"bboxes": [{"bbox": [xmin1, ymin1, xmax1, ymax1], "class": "cls1"},{"bbox": [xmin2, ymin2, xmax2, ymax2], "class": "cls2"}]},"image_name_2": {"bboxes": [{"bbox": [xmin3, ymin3, xmax3, ymax3], "class": "cls1"},{"bbox": [xmin4, ymin4, xmax4, ymax4], "class": "cls2"},{"bbox": [xmin5, ymin5, xmax5, ymax5], "class": "cls3"}]}}
import os,jsonfrom PIL import Imageimg_data_dict={}path = os.getcwd()inputPar = os.path.join(path, r'Dataset')folders = os.listdir(inputPar)clsfile = os.path.join(path, 'classes.txt')with open(clsfile) as tf:clsnames = [cl.strip() for cl in tf.readlines()]for folder in folders:if folder in ["test"]:inputChild = os.path.join(inputPar, folder,"images")files = os.listdir(inputChild)for file in files:imgpath=os.path.join(inputChild,file)img_data_dict[file]=[]fileName, ext = os.path.splitext(file)finput = os.path.join(inputPar,folder,"labels", fileName + '.txt')with open(finput) as tf:Yolodata = tf.readlines()if os.path.exists(imgpath):print("plotting >>",fileName + '.jpg')img = Image.open(imgpath)width, height = img.sizefor obj in Yolodata:clsName = clsnames[int(obj.split(' ')[0])]xnew = float(obj.split(' ')[1])ynew = float(obj.split(' ')[2])wnew = float(obj.split(' ')[3])hnew = float(obj.split(' ')[4])# box sizedw = 1 / widthdh = 1 / height# coordinatesxmax = int(((2 * xnew) + wnew) / (2 * dw))xmin = int(((2 * xnew) - wnew) / (2 * dw))ymax = int(((2 * ynew) + hnew) / (2 * dh))ymin = int(((2 * ynew) - hnew) / (2 * dh))bbx_dict={"Bbox":[xmin,ymin,xmax,ymax],"class":f"{clsName}"}if file in img_data_dict:img_data_dict[file].append(bbx_dict)else:print(f'{imgpath} >> img not found:')with open("img_gt.json","w") as f:json.dump(img_data_dict,f,indent=4)
import cv2,json,osfrom ultralytics import YOLOonnx_model = YOLO("best.onnx")class_names=onnx_model.namesimg_data_dict={}path = os.getcwd()inputPar = os.path.join(path, r'Dataset')folders = os.listdir(inputPar)for folder in folders:if folder in ["test"]:inputChild = os.path.join(inputPar, folder,"images")files = os.listdir(inputChild)for file in files:img_data_dict[file]=[]imgpath=os.path.join(inputChild,file)image = cv2.imread(imgpath)# Run inferenceresults = onnx_model(image)# Extract predictionsfor result in results:boxes = result.boxes # get bounding boxesfor box in boxes:x1, y1, x2, y2 = map(int, box.xyxy[0].tolist()) # Bounding box coordinatesconf = box.conf.item() # Confidence scoreclass_id = int(box.cls.item()) # Class IDclsName=class_names[class_id]bbx_dict={"Bbox":[x1, y1, x2, y2],"class":f"{clsName}"}if file in img_data_dict:img_data_dict[file].append(bbx_dict)with open("img_pred.json","w") as f:json.dump(img_data_dict,f,indent=4)
import jsondef load_data(ground_truth_file, predictions_file):with open(ground_truth_file) as f:ground_truth = json.load(f)with open(predictions_file) as f:predictions = json.load(f)return ground_truth, predictions# Load dataground_truth_file = 'img_gt.json'predictions_file = 'img_pred.json'ground_truth, predictions = load_data(ground_truth_file, predictions_file)# Make a copy of the dataground_truth_upd, predictions_upd = ground_truth.copy(), predictions.copy()# Update the lists so they have the same lengthfor gt_key, pred_key in zip(ground_truth, predictions):# Get the annotations for the current keygt_annotations = ground_truth[gt_key]pred_annotations = predictions[pred_key]if len(gt_annotations) != len(pred_annotations):gt_len = len(gt_annotations)pred_len = len(pred_annotations)# Add padding to the smaller listif gt_len < pred_len:# Pad ground truth with empty boxes and None classfor _ in range(pred_len - gt_len):ground_truth_upd[gt_key].append({"Bbox": [0, 0, 0, 0], "class": None})elif pred_len < gt_len:# Pad predictions with empty boxes and None classfor _ in range(gt_len - pred_len):predictions_upd[pred_key].append({"Bbox": [0, 0, 0, 0], "class": None})# Save updated datawith open("img_gt_upd.json", "w") as f:json.dump(ground_truth_upd, f, indent=4)with open("img_pred_upd.json", "w") as f:json.dump(predictions_upd, f, indent=4)
import jsonimport numpy as npimport matplotlib.pyplot as pltfrom sklearn.metrics import classification_reportdef load_data(ground_truth_file, predictions_file):with open(ground_truth_file) as f:ground_truth = json.load(f)with open(predictions_file) as f:predictions = json.load(f)return ground_truth, predictionsdef iou(box1, box2):# Calculate the intersection coordinatesxi1 = max(box1[0], box2[0])yi1 = max(box1[1], box2[1])xi2 = min(box1[2], box2[2])yi2 = min(box1[3], box2[3])# Calculate the area of intersectionintersection_area = max(0, xi2 - xi1) * max(0, yi2 - yi1)# Calculate the area of both boxesbox1_area = (box1[2] - box1[0]) * (box1[3] - box1[1])box2_area = (box2[2] - box2[0]) * (box2[3] - box2[1])# Calculate the area of unionunion_area = box1_area + box2_area - intersection_areareturn intersection_area / union_area if union_area > 0 else 0def calculate_metrics(ground_truth, predictions, iou_threshold=0.5):tp = 0 # True Positivesfp = 0 # False Positivesfn = 0 # False Negatives# For classification reporty_true = [] # Ground truth classesy_pred = [] # Predicted classesfor image in ground_truth:gt_boxes = ground_truth[image]pred_boxes = predictions[image]matched_gt = [False] * len(gt_boxes) # Track which ground truths have been matchedfor pred in pred_boxes:pred_box = pred['Bbox']pred_class = pred['class']# Append predicted class for reportbest_iou = 0best_gt_idx = -1for idx, gt in enumerate(gt_boxes):gt_box = gt['Bbox']gt_class = gt['class']# Only consider ground truths that match the predicted classif gt_class == pred_class and not matched_gt[idx]:current_iou = iou(pred_box, gt_box)if current_iou > best_iou:best_iou = current_ioubest_gt_idx = idx# Check if the best IoU exceeds the thresholdif best_iou >= iou_threshold and best_gt_idx != -1:tp += 1 # Count as true positivematched_gt[best_gt_idx] = True # Mark this ground truth as matchedelse:fp += 1 # Count as false positivefn += matched_gt.count(False) # Count unmatched ground truths as false negatives# Append ground truth classes for the reporty_true.extend(gt['class'] for gt in gt_boxes)y_pred.extend(pred['class'] for pred in pred_boxes)y_true = [label if label is not None else -1 for label in y_true]y_pred = [label if label is not None else -1 for label in y_pred]# Calculate precision, recall, F1 scoreprecision = tp / (tp + fp) if (tp + fp) > 0 else 0recall = tp / (tp + fn) if (tp + fn) > 0 else 0f1_score = 2 * (precision * recall) / (precision + recall) if (precision + recall) > 0 else 0return tp, fp, fn, precision, recall, f1_score, y_true, y_preddef plot_metrics(precision, recall, f1_score):metrics = [precision, recall, f1_score]labels = ['Precision', 'Recall', 'F1 Score']plt.figure(figsize=(8, 5))plt.bar(labels, metrics, color=['blue', 'orange', 'green'])plt.ylim(0, 1)plt.ylabel('Score')plt.title('Performance Metrics')plt.grid(axis='y')for i, v in enumerate(metrics):plt.text(i, v + 0.02, f"{v:.2f}", ha='center', va='bottom')plt.show()def main(ground_truth_file, predictions_file):ground_truth, predictions = load_data(ground_truth_file, predictions_file)tp, fp, fn, precision, recall, f1_score, y_true, y_pred = calculate_metrics(ground_truth, predictions)print(f"True Positives: {tp}")print(f"False Positives: {fp}")print(f"False Negatives: {fn}")print(f"Precision: {precision:.2f}")print(f"Recall: {recall:.2f}")print(f"F1 Score: {f1_score:.2f}")# Generate classification reportprint("\nClassification Report:")print(f"Length of y_true: {len(y_true)}")print(f"Length of y_pred: {len(y_pred)}")print(classification_report(y_true, y_pred))# Plot metricsplot_metrics(precision, recall, f1_score)# Example usageground_truth_file = 'img_gt_upd.json'predictions_file = 'img_pred_upd.json'main(ground_truth_file, predictions_file)
—THE END—
下载1：Pytorch常用函数手册
欢迎加入CV学习交流微信群！
