检测与标注

Supervision 提供了一个无缝的流程，用于标注由各种物体检测和分割模型生成的预测结果。本指南展示了如何使用 Inference、Ultralytics 或 Transformers 等包进行推理。接着，你将学习如何将这些预测导入 Supervision 并用于标注原始图像。

基本标注示例

运行检测

首先，你需要从你的物体检测或分割模型中获取预测结果。

“Inference”

  1. import cv2
  2. from inference import get_model
  4. model = get_model(model_id="yolov8n-640")
  5. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  6. results = model.infer(image)[0]

“Ultralytics”

  1. import cv2
  2. from ultralytics import YOLO
  4. model = YOLO("yolov8n.pt")
  5. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  6. results = model(image)[0]

“Transformers”

  1. import torch
  2. from PIL import Image
  3. from transformers import DetrImageProcessor, DetrForObjectDetection
  5. processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
  6. model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")
  8. image = Image.open(<SOURCE_IMAGE_PATH>)
  9. inputs = processor(images=image, return_tensors="pt")
  11. with torch.no_grad():
  12.     outputs = model(**inputs)
  14. width, height = image.size
  15. target_size = torch.tensor([[height, width]])
  16. results = processor.post_process_object_detection(
  17.     outputs=outputs, target_sizes=target_size)[0]

将预测结果加载到 Supervision

现在我们有了模型的预测结果，可以将它们加载到 Supervision 中。

“Inference”

我们可以使用 sv.Detections.from_inference 方法加载，该方法支持检测和分割模型的结果。

  1. import cv2
  2. import supervision as sv
  3. from inference import get_model
  5. model = get_model(model_id="yolov8n-640")
  6. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  7. results = model.infer(image)[0]
  8. detections = sv.Detections.from_inference(results)

“Ultralytics”

使用 sv.Detections.from_ultralytics 方法加载，同样支持检测和分割模型结果。

  1. import cv2
  2. import supervision as sv
  3. from ultralytics import YOLO
  5. model = YOLO("yolov8n.pt")
  6. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  7. results = model(image)[0]
  8. detections = sv.Detections.from_ultralytics(results)

“Transformers”

使用 sv.Detections.from_transformers 方法加载，支持检测和分割模型结果。

  1. import torch
  2. import supervision as sv
  3. from PIL import Image
  4. from transformers import DetrImageProcessor, DetrForObjectDetection
  6. processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
  7. model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")
  9. image = Image.open(<SOURCE_IMAGE_PATH>)
  10. inputs = processor(images=image, return_tensors="pt")
  12. with torch.no_grad():
  13.     outputs = model(**inputs)
  15. width, height = image.size
  16. target_size = torch.tensor([[height, width]])
  17. results = processor.post_process_object_detection(
  18.     outputs=outputs, target_sizes=target_size)[0]
  19. detections = sv.Detections.from_transformers(
  20.     transformers_results=results,
  21.     id2label=model.config.id2label)

你也可以使用以下方法加载来自其他计算机视觉框架和库的预测：

使用检测结果标注图像

最后，我们用预测结果对图像进行标注。因为我们使用的是物体检测模型，我们将用到 sv.BoxAnnotator 和 sv.LabelAnnotator 这两个类。

“Inference”

  1. import cv2
  2. import supervision as sv
  3. from inference import get_model
  5. model = get_model(model_id="yolov8n-640")
  6. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  7. results = model.infer(image)[0]
  8. detections = sv.Detections.from_inference(results)
  10. box_annotator = sv.BoxAnnotator()
  11. label_annotator = sv.LabelAnnotator()
  13. annotated_image = box_annotator.annotate(
  14.     scene=image, detections=detections)
  15. annotated_image = label_annotator.annotate(
  16.     scene=annotated_image, detections=detections)

“Ultralytics”

  1. import cv2
  2. import supervision as sv
  3. from ultralytics import YOLO
  5. model = YOLO("yolov8n.pt")
  6. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  7. results = model(image)[0]
  8. detections = sv.Detections.from_ultralytics(results)
  10. box_annotator = sv.BoxAnnotator()
  11. label_annotator = sv.LabelAnnotator()
  13. annotated_image = box_annotator.annotate(
  14.     scene=image, detections=detections)
  15. annotated_image = label_annotator.annotate(
  16.     scene=annotated_image, detections=detections)

“Transformers”

  1. import torch
  2. import supervision as sv
  3. from PIL import Image
  4. from transformers import DetrImageProcessor, DetrForObjectDetection
  6. processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
  7. model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")
  9. image = Image.open(<SOURCE_IMAGE_PATH>)
  10. inputs = processor(images=image, return_tensors="pt")
  12. with torch.no_grad():
  13.     outputs = model(**inputs)
  15. width, height = image.size
  16. target_size = torch.tensor([[height, width]])
  17. results = processor.post_process_object_detection(
  18.     outputs=outputs, target_sizes=target_size)[0]
  19. detections = sv.Detections.from_transformers(
  20.     transformers_results=results,
  21.     id2label=model.config.id2label)
  23. box_annotator = sv.BoxAnnotator()
  24. label_annotator = sv.LabelAnnotator()
  26. annotated_image = box_annotator.annotate(
  27.     scene=image, detections=detections)
  28. annotated_image = label_annotator.annotate(
  29.     scene=annotated_image, detections=detections)

基本标注示例

显示自定义标签

默认情况下，sv.LabelAnnotator 会使用检测结果的 class_name（如果有）或 class_id 作为标签。你可以通过向 annotate 方法传入自定义的 labels 列表来覆盖这一行为。

“Inference”

  1. import cv2
  2. import supervision as sv
  3. from inference import get_model
  5. model = get_model(model_id="yolov8n-640")
  6. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  7. results = model.infer(image)[0]
  8. detections = sv.Detections.from_inference(results)
  10. box_annotator = sv.BoxAnnotator()
  11. label_annotator = sv.LabelAnnotator()
  13. labels = [
  14.     f"{class_name} {confidence:.2f}"
  15.     for class_name, confidence
  16.     in zip(detections['class_name'], detections.confidence)
  17. ]
  19. annotated_image = box_annotator.annotate(
  20.     scene=image, detections=detections)
  21. annotated_image = label_annotator.annotate(
  22.     scene=annotated_image, detections=detections, labels=labels)

“Ultralytics”

  1. import cv2
  2. import supervision as sv
  3. from ultralytics import YOLO
  5. model = YOLO("yolov8n.pt")
  6. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  7. results = model(image)[0]
  8. detections = sv.Detections.from_ultralytics(results)
  10. box_annotator = sv.BoxAnnotator()
  11. label_annotator = sv.LabelAnnotator()
  13. labels = [
  14.     f"{class_name} {confidence:.2f}"
  15.     for class_name, confidence
  16.     in zip(detections['class_name'], detections.confidence)
  17. ]
  19. annotated_image = box_annotator.annotate(
  20.     scene=image, detections=detections)
  21. annotated_image = label_annotator.annotate(
  22.     scene=annotated_image, detections=detections, labels=labels)

“Transformers”

  1. import torch
  2. import supervision as sv
  3. from PIL import Image
  4. from transformers import DetrImageProcessor, DetrForObjectDetection
  6. processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
  7. model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")
  9. image = Image.open(<SOURCE_IMAGE_PATH>)
  10. inputs = processor(images=image, return_tensors="pt")
  12. with torch.no_grad():
  13.     outputs = model(**inputs)
  15. width, height = image.size
  16. target_size = torch.tensor([[height, width]])
  17. results = processor.post_process_object_detection(
  18.     outputs=outputs, target_sizes=target_size)[0]
  19. detections = sv.Detections.from_transformers(
  20.     transformers_results=results,
  21.     id2label=model.config.id2label)
  23. box_annotator = sv.BoxAnnotator()
  24. label_annotator = sv.LabelAnnotator()
  26. labels = [
  27.     f"{class_name} {confidence:.2f}"
  28.     for class_name, confidence
  29.     in zip(detections['class_name'], detections.confidence)
  30. ]
  32. annotated_image = box_annotator.annotate(
  33.     scene=image, detections=detections)
  34. annotated_image = label_annotator.annotate(
  35.     scene=annotated_image, detections=detections, labels=labels)

自定义标签标注示例

使用分割结果标注图像

如果你使用的是分割模型，sv.MaskAnnotator 可以替代 sv.BoxAnnotator，用来绘制掩码而不是边框。

“Inference”

  1. import cv2
  2. import supervision as sv
  3. from inference import get_model
  5. model = get_model(model_id="yolov8n-seg-640")
  6. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  7. results = model.infer(image)[0]
  8. detections = sv.Detections.from_inference(results)
  10. mask_annotator = sv.MaskAnnotator()
  11. label_annotator = sv.LabelAnnotator(text_position=sv.Position.CENTER_OF_MASS)
  13. annotated_image = mask_annotator.annotate(
  14.     scene=image, detections=detections)
  15. annotated_image = label_annotator.annotate(
  16.     scene=annotated_image, detections=detections)

“Ultralytics”

  1. import cv2
  2. import supervision as sv
  3. from ultralytics import YOLO
  5. model = YOLO("yolov8n-seg.pt")
  6. image = cv2.imread(<SOURCE_IMAGE_PATH>)
  7. results = model(image)[0]
  8. detections = sv.Detections.from_ultralytics(results)
  10. mask_annotator = sv.MaskAnnotator()
  11. label_annotator = sv.LabelAnnotator(text_position=sv.Position.CENTER_OF_MASS)
  13. annotated_image = mask_annotator.annotate(
  14.     scene=image, detections=detections)
  15. annotated_image = label_annotator.annotate(
  16.     scene=annotated_image, detections=detections)

“Transformers”

  1. import torch
  2. import supervision as sv
  3. from PIL import Image
  4. from transformers import DetrImageProcessor, DetrForSegmentation
  6. processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50-panoptic")
  7. model = DetrForSegmentation.from_pretrained("facebook/detr-resnet-50-panoptic")
  9. image = Image.open(<SOURCE_IMAGE_PATH>)
  10. inputs = processor(images=image, return_tensors="pt")
  12. with torch.no_grad():
  13.     outputs = model(**inputs)
  15. width, height = image.size
  16. target_size = torch.tensor([[height, width]])
  17. results = processor.post_process_segmentation(
  18.     outputs=outputs, target_sizes=target_size)[0]
  19. detections = sv.Detections.from_transformers(
  20.     transformers_results=results,
  21.     id2label=model.config.id2label)
  23. mask_annotator = sv.MaskAnnotator()
  24. label_annotator = sv.LabelAnnotator(text_position=sv.Position.CENTER_OF_MASS)
  26. labels = [
  27.     f"{class_name} {confidence:.2f}"
  28.     for class_name, confidence
  29.     in zip(detections['class_name'], detections.confidence)
  30. ]
  32. annotated_image = mask_annotator.annotate(
  33.     scene=image, detections=detections)
  34. annotated_image = label_annotator.annotate(
  35.     scene=annotated_image, detections=detections, labels=labels)

分割标注示例