4. 🚀YOLO12: Attention-Centric Object Detection

4.1 Overview

Newly released YOLOv12 uses a new attention-based design instead of the traditional CNN approach while keeping its fast detection speed. It improves accuracy with advanced attention mechanisms and a better network structure, making it one of the most powerful real-time object detection models.

4.2 YOLOv12: Key Innovations

Area Attention Mechanismp

A new way for the model to focus on important areas while keeping calculations fast. It splits images into smaller parts (default: 4) to process them efficiently, reducing computation compared to standard attention methods.

Residual Efficient Layer Aggregation Networks (R-ELAN)

An upgraded method for combining features, improving learning in large attention-based models. It uses special connections (residual scaling) and a bottleneck-like structure for better optimization.

Optimized Attention Architecture

YOLOv12 fine-tunes attention mechanisms for efficiency, using:

• FlashAttention to reduce memory use.
• No positional encoding for faster processing.
• A better balance between attention and feed-forward layers.
• Fewer layers for easier optimization.
• 7x7 separable convolution to help understand positions without extra complexity.

Supports Multiple Tasks

Works for object detection, segmentation, image classification, pose estimation, and detecting rotated objects.

Higher Accuracy with Fewer Parameters

Achieves better results while using less computational power, making it both fast and efficient.

Flexible for Different Devices

Can run on anything from small edge devices to powerful cloud servers.

4.3 Supported Tasks and Modes

YOLO12 supports a variety of computer vision tasks. The table below shows task support and the operational modes (Inference, Validation, Training, and Export) enabled for each:

Model Type	Task	Inference	Validation	Training	Export
YOLO12	Detection	✅	✅	✅	✅
YOLO12-seg	Segmentation	✅	✅	✅	✅
YOLO12-pose	Pose	✅	✅	✅	✅
YOLO12-cls	Classification	✅	✅	✅	✅
YOLO12-obb	OBB	✅	✅	✅	✅

4.4 Performance Metrics

YOLO12 demonstrates significant accuracy improvements across all model scales, with some trade-offs in speed compared to the fastest prior YOLO models. Below are quantitative results for object detection on the COCO validation dataset:

The table shows that YOLO12 models are pre-trained on the COCO dataset and include different versions: YOLO12n.pt, YOLO12s.pt, YOLO12m.pt, YOLO12l.pt, and YOLO12x.pt.

• Inference speed measured on an NVIDIA T4 GPU with TensorRT FP16 precision.
• YOLO12n.pt is the smallest and fastest model but has lower accuracy.
• YOLO12x.pt is the largest model with the highest accuracy but requires significantly more computational power. It is much slower due to its complexity.
• These models are not optimized for normal PC-based CPUs. Running them effectively requires a GPU.

For our demo exercises, we will use NVIDIA’s T4 GPU on Google Colab Notebook for efficient performance.

5. 🚀The Role of COCO Dataset in Object Detection

5.1 COCO Dataset Overview

The COCO (Common Objects in Context) dataset is a large-scale dataset designed for object detection, segmentation, and image captioning. It is widely used in computer vision tasks, particularly in deep learning models for object recognition and instance segmentation.

YOLO uses COCO dataset by default.

Key Features of COCO Dataset:

• Contains 330,000 images (about 200,000 labeled images).
• Includes 1.5 million object instances.
• Provides 80 object categories for detection and segmentation.
• Supports 5 captions per image for image captioning tasks.
• Includes stuff annotations with 91 stuff classes (e.g., sky, road, grass).
• Offers keypoint annotations for person pose estimation with 17 keypoints per person.

How Many Classes Are in COCO?

COCO has 80 object classes for detection and segmentation, plus 91 stuff classes. Below is a list of the 80 common object categories in COCO:

COCO Dataset Object Classes

5.2 List of COCO Dataset Object Classes:

Default pre-trained COCO Dataset models contain following objects for detections.

1. Person 2. Bicycle 3. Car 4. Motorcycle 5. Airplane 6. Bus 7. Train 8. Truck 9. Boat 10. Traffic light 11. Fire hydrant 12. Stop sign 13. Parking meter 14. Bench 15. Bird 16. Cat

17. Dog 18. Horse 19. Sheep 20. Cow 21. Elephant 22. Bear 23. Zebra 24. Giraffe 25. Backpack 26. Umbrella 27. Handbag 28. Tie 29. Suitcase 30. Frisbee 31. Skis 32. Snowboard

33. Sports ball 34. Kite 35. Baseball bat 36. Baseball glove 37. Skateboard 38. Surfboard 39. Tennis racket 40. Bottle 41. Wine glass 42. Cup 43. Fork 44. Knife 45. Spoon 46. Bowl 47. Banana 48. Apple

49. Sandwich 50. Orange 51. Broccoli 52. Carrot 53. Hot dog 54. Pizza 55. Donut 56. Cake 57. Chair 58. Couch 59. Potted plant 60. Bed 61. Dining table 62. Toilet 63. TV 64. Laptop

65. Mouse 66. Remote 67. Keyboard 68. Cell phone 69. Microwave 70. Oven 71. Toaster 72. Sink 73. Refrigerator 74. Book 75. Clock 76. Vase 77. Scissors 78. Teddy bear 79. Hairdryer 80. Toothbrush

• We can detect any of the above listed items using YOLO12n.pt, YOLO12s.pt, YOLO12m.pt, YOLO12l.pt and YOLO12x.pt pre-trained models.
• But to detect other items like forest fire, brain tumer, car registration license plates that are not listed in Coco dataset, we need to create pre-trained custom dataset.
• In this project we shall use default pretrained model YOLO12n.pt pretrained model to detect persons, cars, trucks, bicycles and so on.
• In our next project we shall create customised pre-trained models to detect forest fire, brain tumer, car registration license plates

6. What are Bounding Boxes?

🚀6.1 Bounding Boxes Overview

Bounding boxes are rectangular boxes used in object detection to define the location of an object in an image. They enclose the detected objects and provide coordinates (x, y, width, height).

Why and How Does YOLO Create Bounding Boxes?

YOLO (You Only Look Once) generates bounding boxes to detect objects in real-time. It does this by:

• Dividing an image into a grid.
• Assigning each grid cell responsibility for predicting objects.
• Using anchor boxes to estimate object positions and dimensions.
• Applying non-maximum suppression (NMS) to remove duplicate detections.

6.2 What Are .pt and .yaml Files?

.pt Files

A .pt file is a PyTorch model checkpoint that stores a trained YOLO model. It contains:

• Model architecture.
• Trained weights.
• Optimizer states (if saved).

Example: yolov12.pt, yolov12s.pt

These files are created using:

torch.save(model.state_dict(), "model.pt")

.yaml Files

A .yaml file is a configuration file that defines model architecture, dataset details, and training settings. It is used to train and fine-tune YOLO models.

It contains:

• Number of object classes.
• Paths to training and validation datasets.
• Model layers and parameters.

Example: yolov12.yaml, coco.yaml

Sample content of a .yaml file:

nc: 80
train: ./data/train/images
val: ./data/val/images
names: ['person', 'bicycle', 'car', ...]

How Are These Files Used?

• .pt files are used for model inference and deployment.
• .yaml files configure training and dataset paths.
• YOLO loads both files when training or running object detection.

7. Object Detection demo using YOLO12n default dataset

7.1 Practical demo

Demo: Using YOLOv12 Pre-trained Model (YOLO12n.pt)

Step 1: Login to Google Drive

If you don’t have a Google Drive account, create one and log in.

Step 2: Upload YOLOV12 Folder

Ensure your YOLOV12 folder on your local computer contains the following images and videos:

• football1.jpg
• soccer1b.mp4
• car2.mp4

Upload this folder to your Google Drive.

Step 3: Open Google Colab

Go to Google Colab and log in.

Step 4: Configure Hardware (T4 GPU)

1. Click Runtime in the top menu.
2. Click Change Runtime Type.
3. Select T4 GPU as the hardware accelerator.
4. Click Save.

Step 5: Verify GPU Setup

Run the following command in a new cell:

import tensorflow as tf
    

print("Num GPUs Available:", len(tf.config.list_physical_devices('GPU')))

If the output is Num GPUs Available: 1, then the GPU is configured correctly.

Step 6: Install Ultralytics

Similar to Command Prompt, "pip install" command can be executed on Colab Notebook cell. But in Notebook cell '!' should be added before pip install command like "!pip install ultralytics" as shown below:

!pip install ultralytics

Click the "Run" button. Once installed, hide the output and add a new cell.

Note: In Command Prompt '!' should not be added. To install Ultralytics in Command Prompt we shall have to use 'pip install ultralytics'. In Jupyter Notebook cell also, we can install Ultralytics using '!' before the pip command.

Step 7: Mount Google Drive

Run the following command in a new Colab cell:

from google.colab import drive
drive.mount('/content/drive')

Step 8: Access YOLO Model

You can either download YOLO12n.pt and upload it to your Google Drive or directly load it from Ultralytics.

Step 9: Change to YOLOV12 Directory

Run this command to go to the mounted drive:

%cd /content/drive/MyDrive/YOLOV12

Step 10: Load YOLO Model

Run the following command in a new cell:

from ultralytics import YOLO

Then, load the pre-trained model:

model = YOLO('yolo12n.pt')

Step 11: Run Object Detection on Images

Run the following command to detect objects in an image:

result = model('/content/drive/MyDrive/YOLOV12/football1.jpg', save=True)

The detected image will be saved in the folder: runs/detect/predict.

Step 12: Run Object Detection on Videos

Run similar commands for video files:

result = model('/content/drive/MyDrive/YOLOV12/soccer1b.mp4', save=True)
result = model('/content/drive/MyDrive/YOLOV12/car2.mp4', save=True)

Detected and image and videos will be saved in the runs/detect/predict folder as:

• football1.jpg
• soccer1b.avi
• car2.avi

Download the image and video files.

Step 13: Convert AVI to WebM

we can convert the .avi files to .webm format for web playback using online tools:

• CloudConvert
• EZGIF

Step 14: Test how detections are successful:

Double click on football1.jpg

YOLOv12 Object Detection Comparison

On the right hand side, image showing that players are detected with 90% confidence and the ball is detected with 89% confidence. Players and the ball are encicled by bounding boxes after detection.

Video Example 1: Before and After Detections

Video before detection

VIDEO: Click on play arrow icon to play the video.

Video After detection

VIDEO: Click on play arrow icon to play the video.

Video Example 2: Before and After Detections

Video before detection

VIDEO: Click on play arrow icon to play the video.

Video After detection

VIDEO: Click on play arrow icon to play the video.

Final Thoughts

By following these steps, you can use YOLOv12's pre-trained model for object detection on images and videos. The detected outputs can be saved and converted into web-friendly formats for easy sharing.

8. Conclusion

In this demonstration, we successfully utilized the YOLOv12n.pt pre-trained model to detect objects in both images and videos. Using Google Colab with a T4 GPU, we performed object detection on sample media files with the Ultralytics YOLO framework. The detected objects were accurately identified and highlighted using bounding boxes. Additionally, we explored methods to convert and display the results efficiently.

As the next step, we will create a new webpage demonstrating object detection using a custom dataset, allowing us to train YOLO on specific objects for improved accuracy in specialized applications.

9. References

• Redmon, J., & Farhadi, A. (2016). YOLO: Real-Time Object Detection. arXiv preprint arXiv:1506.02640.
• Ultralytics. (2023). YOLOv11 Documentation. Available at: https://ultralytics.com/
• PyTorch Documentation. Available at: https://pytorch.org/
• COCO Dataset:
• Google Colab:
• CloudConvert for video conversion:
• EZGIF for video format conversion: