YOLOv8 P2 and P6 variants for small object detection

[matijasukovic]

Hello, thank you for your time. I am training a model to detect small objects (as small as 10px in width) in large images (1920x1080). I saw in a similar forum post someone recommended using one of the yolov8-p2 models to improve performance on small objects. I tried them and indeed they perform better, but I still have subpar results. Looking at the models list inside of the Ultralytics library, I see there is a P6 variant. I know P2 and P6 relates to feature pyramids in some way, but I found very little information regarding P2 online, and none at all for P6. Maybe I'm looking it up wrong. Can you shed some light on what exactly is different about these models compared to the normal versions? What are the tradeoffs when using these? Will the P6 variants increase performance on small objects further?

[glenn-jocher]

Hello! I'm glad to hear you're exploring the P2 and P6 variants for small object detection. 🕵️‍♂️

The P2 and P6 variants refer to the strides of the feature pyramids used in the models. The P2 models have a stride of 2, meaning they preserve more fine-grained spatial details, which can be beneficial for detecting smaller objects. On the other hand, P6 models have a stride of 6, which makes them more suited for larger objects as they capture higher-level features.

For small object detection, P2 variants are typically more effective due to their finer feature maps. However, using P6 won't necessarily improve small object detection; it might even hinder it because of the coarser feature maps.

The tradeoffs when using these specialized models usually involve a balance between speed and accuracy. P2 models might be slower but better for small objects, while P6 models could be faster but less precise for small details.

Here's a simple example of how you might train a P2 model:

from ultralytics import YOLO

# Load a P2 model
model = YOLO('yolov8n-p2.pt')

# Train the model on your dataset
model.train(data='your_dataset.yaml', epochs=50, imgsz=1920)

Remember to adjust the imgsz parameter to match your input image size for optimal results. Good luck with your training, and let us know if you have any more questions! 🚀

[park-sungho-8]

Hi @glenn-jocher.

When I was following your codes, I encountered an issue that I can't find anywhere to download 'yolov8n-p2.pt'.

Would you share a pt file or link to find it please.

[glenn-jocher]

@park-sungho-8 there are no pretrained models, but you can construct one from the provided YAML, i.e. model=YOLO('yolov8n-p2.yaml')

[abdulmanafpv]

Hi @glenn-jocher
I have 4 doubts will you please help me to clarify my doubts

1. I have a doubt how we can customise the yolov8-p2.yaml for detecting more small objects for example P2/4 how we can change to P2/2 , P3/8 as P3/4 , p4/16 as P4/8, p5/32 as p5/16 so this will help me to detect the xx-small , x- small, small and medium.

2. Also can we change any one of these for example can I change P2/4 as P2/2 without change any other variants, if we do this like there have any problem in our model accuracy.

3. when I change the yaml file in custom way how we can use yolov8s -p2.yaml during training time. Consider if I change the yaml file name in to any other name like custom-p2.yaml so how I can get the small model architecture during training time. What would be the training command for this. ?

4. can I change the scale factor of 2 into 4 in upsample part of head in P2 variant is there any problem when I change into 4.

• [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  • [[-1, 4], 1, Concat, [1]] # cat backbone P3
  • [-1, 3, C2f, [256]] # 15 (P3/8-small)

[glenn-jocher]

@abdulmanafpv to change P2/4 to P2/2 in the YAML file, adjust the stride in the model's backbone section to reduce the downsampling rate. Make sure to test the model for performance impacts.

[kirimaru-jp]

@glenn-jocher Thank you for the great info!
Is it a good idea to train YOLOv8 P2 using both small and large objects simultaneously, since I'd like to detect objects with different sizes?

[glenn-jocher]

Training YOLOv8 P2 on both small and large objects can be effective, as it helps the model generalize across different object sizes. Just ensure your dataset is well-balanced.

[zeobec]

@glenn-jocher hii i have a question regarding on the model parameters. From my understanding, p2-yolov8 adds extra components compared to yolov8. Logically, the parameters will increase right? Based on my observation during training, I noticed that parameters for p2-yolov8n was much lesser than yolov8n. May i know if anyone obtained the same results like me and I wonder why that thing happened

[glenn-jocher]

@zeobec the P2 variant of YOLOv8 is designed to optimize for small object detection, which may involve architectural changes that don't necessarily increase parameter count. The reduction in parameters could be due to efficient design choices that maintain performance while using fewer resources.