Object Cropping using Ultralytics YOLO26
What is Object Cropping?#
Object cropping with Ultralytics YOLO26 involves isolating and extracting specific detected objects from an image or video. YOLO26 locates each object with a bounding box, and the cropper saves that region as a separate image for further analysis or processing.
Watch: Object Cropping using Ultralytics YOLO
Advantages of Object Cropping#
- Focused Analysis: YOLO26 enables targeted object cropping, allowing in-depth examination or processing of individual items within a scene.
- Reduced Data Volume: By extracting only relevant objects, object cropping minimizes data size, making storage, transmission, and subsequent computational tasks more efficient.
- Enhanced Precision: YOLO26's object detection accuracy produces tight crops that preserve the visual detail needed for further analysis.
Visuals#
| Airport Luggage |
|---|
![]() |
| Suitcases Cropping at airport conveyor belt using Ultralytics YOLO26 |
Crop Objects with YOLO26#
Pass your video to the ObjectCropper solution and it detects objects each frame and saves every detection as a separate image. Restrict cropping to specific classes, raise conf to keep only confident detections, and set crop_dir to choose where the crops are written.
# Crop the objects
yolo solutions crop
# Pass a source video
yolo solutions crop source="path/to/video.mp4"
# Crop specific classes
yolo solutions crop classes="[0, 2]"ObjectCropper Arguments#
Here's a table with the ObjectCropper arguments:
| Argument | Type | Default | Description |
|---|---|---|---|
model | str | None | Path to an Ultralytics YOLO model file. |
crop_dir | str | 'cropped-detections' | Directory name for storing cropped detections. |
verbose | bool | True | Enables the solution's per-frame log of input shape, class counts and processing speed. The tracking call itself is always silent. |
The ObjectCropper runs predict rather than track, so it supports the following inference parameters except tracker:
| Argument | Type | Default | Description |
|---|---|---|---|
tracker | str | 'tracktrack.yaml' | Specifies the tracking algorithm to use. Built-in options: botsort.yaml, bytetrack.yaml, ocsort.yaml, deepocsort.yaml, fasttrack.yaml, tracktrack.yaml. |
conf | float | 0.25 | Sets the confidence threshold for detections; lower values allow more objects to be tracked but may include false positives. |
iou | float | 0.5 | Sets the Intersection over Union (IoU) threshold for filtering overlapping detections. |
classes | list | None | Filters results by class index. For example, classes=[0, 2, 3] only tracks the specified classes. |
device | str | None | Specifies the device for inference (e.g., cpu, cuda:0 or 0). Allows users to select between CPU, a specific GPU, or other compute devices for model execution. |
imgsz | int or tuple | 640 | Target inference image size: an integer for a square or a (height, width) pair. Shapes are adjusted to the model stride. |
max_det | int | 300 | Maximum number of detections kept per frame before tracking, limiting the objects a solution processes in dense scenes. |
quantize | int or str | None | Inference precision: 16 (FP16) or 32/unset (FP32) for PyTorch and TorchScript models off CPU; other formats use the precision their artifact and runtime select. Replaces the deprecated half arg. |
FAQ#
Object cropping using Ultralytics YOLO26 involves isolating and extracting specific objects from an image or video based on YOLO26's detection capabilities. This process allows for focused analysis and reduced data volume by using YOLO26 to identify objects accurately and crop them accordingly.
Ultralytics YOLO26 combines accurate detection, fast inference, and a simple API, so cropping takes only a few lines of code and supports focused analysis. YOLO26 also exports to formats such as OpenVINO and TensorRT for real-time deployments on diverse hardware.
By using Ultralytics YOLO26 to crop only relevant objects from your images or videos, you can significantly reduce the data size, making it more efficient for storage and processing. Use a pretrained model, or train one to detect the objects you need, then crop and save only those regions.
Yes, Ultralytics YOLO26 can process real-time video feeds to detect and crop objects dynamically. Its fast inference suits real-time applications such as surveillance, sports analysis, and automated inspection systems.
Ultralytics YOLO26 is optimized for both CPU and GPU environments, but to achieve optimal performance, especially for real-time or high-volume inference, a dedicated GPU (e.g., an NVIDIA RTX or data center GPU) is recommended. For deployment on lightweight devices, consider using CoreML for iOS or LiteRT for Android.
Comments15
Ask a question, share your results, or help someone else.
NOSPAA21
Can i use it so as to train a model especially for printed circuit board defect detection to find defects in certain regions of a pcb?Would this be helpful?
RizwanMunawar
@NOSPAA21 Yes, you can train the Ultralytics YOLO11 model on the circuit board defect detection dataset by following the training documentation. Once the model is trained, you can use the object-cropping solution to crop the different defects that the model detects.
AdityaJaiswal99
Can we make the model, so it does not take the same pictures in the frame again?
Glenn JocherUltralytics@AdityaJaiswal99 thank you for your question! If you're looking to avoid capturing the same objects multiple times within a frame, you can leverage the Non-Maximum Suppression (NMS) feature of YOLOv8. NMS helps in eliminating redundant bounding boxes by keeping only the highest confidence detections and discarding others that overlap significantly.
Here's a brief example of how you can adjust the
iouthreshold to refine your detections and reduce duplicates:from ultralytics import YOLO # Load the model model = YOLO("yolov8n.pt") # Set the IoU threshold to a lower value to reduce overlapping detections results = model.predict("path/to/your/image_or_video", iou=0.5) # Process the results as needed for result in results: boxes = result.boxes.xyxy.cpu().tolist() clss = result.boxes.cls.cpu().tolist() # Further processing...In this example, setting the
iouparameter to a lower value (e.g., 0.5) helps in reducing the number of overlapping boxes, thus minimizing the chances of capturing the same object multiple times.For a more comprehensive guide on object cropping and additional tips, please visit our Object Cropping Guide.
If you encounter any issues or have further questions, feel free to provide a reproducible example, and ensure you are using the latest version of the packages. This will help us assist you more effectively. 😊
Happy coding!
AdityaJaiswal99
How can we change the frame height and width to a specific pixel and change FPS?
Paula DerrengerUltralytics@AdityaJaiswal99 hi there!
To change the frame height, width, and FPS of a video using OpenCV, you can set these properties directly on the
cv2.VideoCaptureandcv2.VideoWriterobjects. Below is an example of how you can achieve this:import cv2 # Load the video cap = cv2.VideoCapture("path/to/video/file.mp4") assert cap.isOpened(), "Error reading video file" # Set the desired frame width, height, and FPS frame_width = 640 # Set your desired width frame_height = 480 # Set your desired height fps = 30 # Set your desired FPS # Create a VideoWriter object to save the modified video video_writer = cv2.VideoWriter("modified_video.avi", cv2.VideoWriter_fourcc(*"mp4v"), fps, (frame_width, frame_height)) while cap.isOpened(): success, frame = cap.read() if not success: break # Resize the frame to the desired dimensions resized_frame = cv2.resize(frame, (frame_width, frame_height)) # Write the resized frame to the output video video_writer.write(resized_frame) # Display the frame (optional) cv2.imshow("Resized Video", resized_frame) if cv2.waitKey(1) & 0xFF == ord("q"): break # Release resources cap.release() video_writer.release() cv2.destroyAllWindows()This script reads a video, resizes each frame to the specified width and height, and writes the modified frames to a new video file with the desired FPS.
For more detailed information on object cropping using YOLOv8, you can refer to our Object Cropping Guide.
Feel free to reach out if you have any more questions! 😊
1
HsnKmlMt
Hi! How can i specify classed to be cropped like for my case i just want people.
Glenn JocherUltralytics(edited)Hello! To crop only people from your video, filter the detections by class. In COCO-trained models 'person' is class 0, so the simplest way is the
classesargument ofpredict:import os import cv2 from ultralytics import YOLO model = YOLO("yolo26n.pt") cap = cv2.VideoCapture("path/to/video/file.mp4") assert cap.isOpened(), "Error reading video file" crop_dir_name = "ultralytics_crop" os.makedirs(crop_dir_name, exist_ok=True) idx = 0 while cap.isOpened(): success, im0 = cap.read() if not success: break results = model.predict(im0, classes=[0]) # 0 = 'person' in COCO boxes = results[0].boxes.xyxy.cpu().tolist() for box in boxes: idx += 1 crop_obj = im0[int(box[1]) : int(box[3]), int(box[0]) : int(box[2])] cv2.imwrite(os.path.join(crop_dir_name, f"person_{idx}.png"), crop_obj) cap.release() cv2.destroyAllWindows()This will crop and save images only for detected persons. Note that
model.namesis a dictionary mapping class IDs to names, so to look up an ID by name usenext(k for k, v in model.names.items() if v == "person"). Adjust according to your specific classes and needs! 👍HsnKmlMt
Thank you, I really appreciate it!
Glenn JocherUltralytics(edited)Hello! We're really glad you found the materials helpful! If you ever need to crop specific classes, such as 'people', filter the detections by class ID. Here's a quick example:
import cv2 from ultralytics import YOLO model = YOLO("yolo26n.pt") # Load your model person_class_id = next(k for k, v in model.names.items() if v == "person") # model.names maps class ID -> name cap = cv2.VideoCapture("path/to/video.mp4") assert cap.isOpened(), "Error reading video file" while cap.isOpened(): ret, frame = cap.read() if not ret: break results = model.predict(frame, classes=[person_class_id]) # only 'person' detections boxes = results[0].boxes.xyxy.cpu().tolist() for idx, box in enumerate(boxes): x1, y1, x2, y2 = map(int, box[:4]) crop_img = frame[y1:y2, x1:x2] cv2.imwrite(f"cropped_person_{idx}.jpg", crop_img)Feel free to replace 'person' with any class you're interested in, and make sure the class name matches what's defined in your model's class names. Let us know if you need further assistance! 🚀
1umairrkhn
Hey Glenn, I am getting this error using your implementation:
person_class_id = model.names.index('person') # Ensure 'person' is in the model's class names ^^^^^^^^^^^^^^^^^ AttributeError: 'dict' object has no attribute 'index'
Glenn JocherUltralytics(edited)Hey there!
It looks like you're trying to use the
.index()method on a dictionary, which isn't supported.model.namesis a dictionary mapping class IDs to class names, not a list. Here's how to find the ID for a class name:model = YOLO("yolo26n.pt") person_class_id = next(k for k, v in model.names.items() if v == "person")For COCO-trained models this is simply
0. If the issue persists, could you share more of your code context? That might help diagnose the issue better. 😊
Glenn JocherUltralytics(edited)Hey there! It looks like you're trying to use the
index()method on a dictionary, which isn't supported.model.namesis a dictionary mapping class IDs to class names, so you need to look up the ID by name:person_class_id = next(k for k, v in model.names.items() if v == "person")Make sure that 'person' is indeed the name of the class you're interested in. Let me know if this helps or if you need further assistance! 😊
zakirh
To those who are facing issue on finding person class:
person_class_id = None for class_id, class_name in model.names.items(): if class_name =='person': person_class_id = class_id break
if person_class_id is None: print("The 'person' class was not found.") else: print(f"'person' class ID: {person_class_id}")
Glenn JocherUltralytics(edited)@zakirh thank you for sharing this helpful approach to finding the class ID for 'person.' For others who might encounter similar issues, note that
model.namesis a dictionary mapping class IDs to class names, so finding an ID by name requires searching the values. If any further clarification is needed, feel free to ask!1
