models/yolo11/

[giscus[bot]]

models/yolo11/

Discover YOLO11, the latest advancement in state-of-the-art object detection, offering unmatched accuracy and efficiency for diverse computer vision tasks.

https://docs.ultralytics.com/models/yolo11/

[UltralyticsAssistant]

👋 Hello, thank you for your interest in Ultralytics and the exciting new YOLO11 model 🚀!

For new users or those curious about YOLO11, we recommend checking out the Docs, which are packed with information on the latest advancements and usage examples.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us diagnose the issue effectively.

If you have a custom training ❓ Question, please include detailed information such as dataset examples and training logs, and ensure you're following our Tips for Best Training Results.

We're excited to have you in the Ultralytics community! If you need real-time support or want to engage in discussions, join us on Discord 🎧. For long-form topics, you might prefer Discourse, or dive into our Subreddit.

Upgrade

Ensure you have the latest ultralytics package with all requirements by running this command in a Python>=3.8 environment with PyTorch>=1.8:

pip install -U ultralytics

Environments

YOLOv8 and, presumably, YOLO11 can be run in these environments, all configured with necessary dependencies:

• Notebooks with free GPU:
• Google Cloud Deep Learning VM - follow the GCP Quickstart Guide.
• Amazon Deep Learning AMI - see the AWS Quickstart Guide.
• Docker Image - details are in the Docker Quickstart Guide

Status

This badge indicates the current status of Ultralytics CI tests, ensuring all YOLO models, including upcoming releases, run smoothly across platforms.

This response was generated automatically 🔄. An Ultralytics engineer will also assist soon. Thank you for your patience and support!

[Pn7Hao]

Hi, My friend and I discovered an issue with the 'In and Out' counting not functioning effectively in count.counter. We then found that the count_objects function was not calculating correctly and fixed it using the dot product. The following code:
.\ultralytics\solutions\object_counter.py

def count_objects(self, track_line, box, track_id, prev_position, cls):
        """
        Helper function to count objects within a polygonal region.

        Args:
            track_line (dict): last 30 frame track record
            box (list): Bounding box data for specific track in current frame
            track_id (int): track ID of the object
            prev_position (tuple): last frame position coordinates of the track
            cls (int): Class index for classwise count updates
        """
        if prev_position is None or track_id in self.counted_ids:
            return

        centroid = self.r_s.centroid
        # dx = (box[0] - prev_position[0]) * (centroid.x - prev_position[0])
        # dy = (box[1] - prev_position[1]) * (centroid.y - prev_position[1])
        
        curr_pos = prev_position
        prev_pos = self.track_history[track_id][-3] if len(self.track_history[track_id]) > 2 else None
        if prev_pos is None:
            return
        
        start_point = self.region[0]
        end_point = self.region[1]
        # A((curr_pos[0]-prev_pos[0]), (curr_pos[1]-prev_pos[1]))
        # B((end_point[0]-start_point[0]), (end_point[1]-start_point[1]))
        d = (curr_pos[0]-prev_pos[0])* (end_point[1]-start_point[1]) - \
            (curr_pos[1]-prev_pos[1]) * (end_point[0]-start_point[0])
        dx = 1
        if d < 0:
            dy = 1
        else:
            dy = -1
        # The remain code

[glenn-jocher]

@Pn7Hao thank you for sharing your solution. If you'd like to contribute your improvements, please consider submitting a pull request to the repository. This way, the community can benefit from your enhancements.

[Pn7Hao]

Hi @glenn-jocher, I have a pull requests already. Thank you for understanding this ❤.

[glenn-jocher]

@Pn7Hao thank you for your pull request. We appreciate your contribution to improving the project. If you have any further questions or need assistance, feel free to ask here.

[Ariya-Kh]

hi
is yolov11 NMS-free like yolov10?

[glenn-jocher]

No, Yolo11 is not NMS-free.

Edited: 13/12/2024

[matpalm]

as of commit #060214d0666258660a018f86874114990873a6a2 that's not how it's currently configured though... (?)

https://github.com/ultralytics/ultralytics/blob/main/ultralytics/cfg/models/11/yolo11.yaml is configured with the Detect head, not the v10Detect head...

so

ultralytics/ultralytics/nn/modules/head.py

Line 21 in 060214d

class Detect(nn.Module):

will have end2end false and only train the one2one head ( as in v8 )

unless there's some other code path enabling of the one2many ?

[glenn-jocher]

The current configuration uses the Detect head, which defaults to one2one mode. If you're looking for one2many functionality, it might require additional modifications not present in the current setup. For further details, please check the latest documentation or code updates.

[matpalm]

which implies it's not NMS-free...

[glenn-jocher]

Thank you for pointing that out. Currently, YOLOv11 uses the Detect head, which includes NMS. If you're looking for NMS-free functionality, it may require further configuration changes. Please check the latest updates or documentation for any modifications.

[Imran-ml]

First of all its great to have another powerful model and congratulations. I have fine tuned and trained yolo11x.pt for my custom objects but I'm facing some issues. My images data resolution is around 6000 x 2000 but while trained the model I set the imgsz=(1280, 640). It work well when I adjust the video stream by zoom in-out and detect all my 60 objects well (objects are small) but when there is some zoom or squeezed then it completely dont detect any object or few only.

Please let me know if I'm missing something or I have to do something else?
I have labelled the data on my original images so resizing making any issues? any ratio issues? Please guide me properly so I can use this power yolo11 model effectively.

[glenn-jocher]

@Imran-ml thank you for your kind words! It seems the issue might be related to the resizing of images during training. Ensure that the aspect ratio is maintained and consider using data augmentation techniques to improve model robustness. Adjusting the imgsz parameter to better match your original image resolution might also help. For further guidance, please refer to the YOLO documentation.

[Skillnoob]

You can try using SAHI with your high resolution images.

[Taufiq-Ahmed]

[Prasoon-Rai]

You need to manually download the model and then use it.

[glenn-jocher]

@Taufiq-Ahmed it seems like the "yolo11n.pt" model file is not available. Please ensure the file path is correct or check if the model is released. For more details, visit the Ultralytics documentation.

[shivakarnati]

@Taufig-Ahmed I had yoloV8 before in my system, so uninstalled the ultralytics package and installed again, now it's working fine with yolo11n.pt model.

[glenn-jocher]

@shivakarnati i'm glad to hear reinstalling the package resolved the issue. If you have any more questions, feel free to ask!

[xuan2261]

Error AMP: checks failed ❌. Anomalies were detected with AMP on your system that may lead to NaN losses or zero-mAP results, so AMP will be disabled during training. "Overriding model.yaml nc=80 with nc=16
Transferred 1009/1015 items from pretrained weights
Freezing layer 'model.23.dfl.conv.weight'
AMP: running Automatic Mixed Precision (AMP) checks with YOLO11n...
Downloading https://github.com/ultralytics/assets/releases/download/v8.3.0/yolo11n.pt to 'yolo11n.pt'...
100%|██████████| 5.35M/5.35M [00:00<00:00, 103MB/s]
AMP: checks failed ❌. Anomalies were detected with AMP on your system that may lead to NaN losses or zero-mAP results, so AMP will be disabled during training."

[glenn-jocher]

@xuan2261 it seems that AMP checks have failed on your system, which may lead to NaN losses or zero-mAP results. As a result, AMP will be disabled during training. For more details on YOLO11, please visit the documentation at https://docs.ultralytics.com/models/yolo11/.

[vishalvijay99]

What is the license for yolov11?

[glenn-jocher]

@vishalvijay99 yOLO11 models are provided under the AGPL-3.0 and Enterprise licenses. For more details, please visit the Ultralytics License page.

[HoangVuAn]

Hi,
Why does YOLO11 have less parameters but takes up more GPU VRAM than YOLOv8?
Is the reason because the calculation is more complicated?

[glenn-jocher]

@HoangVuAn yOLO11's increased GPU VRAM usage, despite having fewer parameters, is due to its enhanced feature extraction and complex architecture, which require more computational resources. For more details, please visit the YOLO11 documentation.

[RayCali]

Then what is the point of the fewer parameters if it still needs more computational resources?

[glenn-jocher]

@RayCali the fewer parameters in YOLO11 contribute to computational efficiency, while its enhanced architecture and feature extraction capabilities require more VRAM to handle complex tasks effectively.

[RayCali]

Ah alright. Could you give me an example of such enhanced architecture or feature extraction capabilities so i could gain some more intuition?

[glenn-jocher]

YOLO11's enhanced architecture includes an improved backbone and neck design, which boosts feature extraction capabilities for more precise object detection. For further insights, please refer to the YOLO11 documentation.

[AISoltani]

Hi, Thanks for your great job!.
How do I apply class weights or handle class imbalance when training a model in Yolov11?

[glenn-jocher]

@AISoltani thank you for your kind words! To handle class imbalance in YOLO11, you can adjust class weights in the training configuration file. This helps balance the influence of each class during training. For more details, please refer to the YOLO11 documentation: https://docs.ultralytics.com/models/yolo11/.

[khanhdu1609]

@glenn-jocher Is it the cfg file that you have mentioned? I also want to fix the class wight too. Thank you very much!

[glenn-jocher]

To apply class weights in YOLO11, you can modify the training configuration by specifying class weights in the dataset configuration YAML file. This helps manage class imbalance by giving different weights to each class during training. For more guidance, please refer to the YOLO11 documentation: YOLO11 Documentation. If you have further questions, feel free to ask!

[Kindstep]

Hi!

Cool to see a new version released! I will soon be doing research (as part of my thesis) on the use of Computer Vision models for physical rehabilitation purposes, and Pose Estimation with accurate keypoint detection will thus be a key factor in the success of my research. Does YOLO11 have integrated keypoint detection?

I am mostly referring to the fact that my predecessor working on a similar project had trouble with YOLOv8, so I am now doing some pre-research on suitable open source models.

[glenn-jocher]

@Kindstep yes, YOLO11 supports keypoint detection with its YOLO11-pose models, making it suitable for tasks like pose estimation in physical rehabilitation. For more details, please visit the YOLO11 documentation.

[AvinashMehta2000]

[glenn-jocher]

@AvinashMehta2000 for YOLO11, you can use any data version from Roboflow that is compatible with YOLO format. Ensure your dataset is properly formatted for YOLOv8 or YOLO11, as both support similar data structures. For more details, refer to the YOLO11 documentation.

[MrYanMYN]

Hey! , Can the model be accelerated using the tensorRT runtime? (Converting the model to .engine etc)

[glenn-jocher]

@MrYanMYN yes, YOLO11 models can be accelerated using TensorRT by exporting them to the .engine format. This can significantly enhance inference speed on compatible NVIDIA GPUs.

[maoxiuyuan]

Hi, thanks for your contribution on YOLO. In YOLOv11, I encountered a question.

I loaded the YOLOv11n model using the following code:
yolo11 = YOLO(r'C:\Users\yolo11\yolo11n.pt')
train_result = yolo11.train(data=r'D:\ultralytics-main\datasets\li_dataset\li_cotton.yaml',
epochs=300, imgsz=800, batch=32, device=[0], workers=0, cache=True)
However, I received an error: AttributeError: Can't get attribute 'C3k2' on <module 'ultralytics.nn.modules.block'>
Looking forward to your reply! 😊

[glenn-jocher]

It seems like the error might be related to a missing or incorrect module. Please ensure your Ultralytics package is up to date. If the issue persists, try reinstalling the package. For further assistance, you can check the Ultralytics documentation.

[maoxiuyuan]

thanks for your reply。I had solved this error.why this error happened is that&nbsp;&nbsp;i use ultralytics-main files which downloaded from yolov8。then i downloaded yolo11 model main file on github again，and the model run successfully 发自我的iPhone

…

------------------ Original ------------------ From:&nbsp;Glenn Jocher ***@***.***&gt; Date:&nbsp;Fri,Oct 4,2024 8:36 PM To:&nbsp;ultralytics/ultralytics ***@***.***&gt; Cc:&nbsp;xiuyuan mao ***@***.***&gt;, Comment ***@***.***&gt; Subject:&nbsp;Re: [ultralytics/ultralytics] models/yolo11/ (Discussion #16541) It seems like the error might be related to a missing or incorrect module. Please ensure your Ultralytics package is up to date. If the issue persists, try reinstalling the package. For further assistance, you can check the&nbsp;Ultralytics documentation. — Reply to this email directly,&nbsp;view it on GitHub, or&nbsp;unsubscribe. You are receiving this because you commented. 发自我的iPhone

------------------ Original ------------------ From: Glenn Jocher ***@***.***&gt; Date: Fri,Oct 4,2024 8:36 PM To: ultralytics/ultralytics ***@***.***&gt; Cc: xiuyuan mao ***@***.***&gt;, Comment ***@***.***&gt; Subject: Re: [ultralytics/ultralytics] models/yolo11/ (Discussion #16541) It seems like the error might be related to a missing or incorrect module. Please ensure your Ultralytics package is up to date. If the issue persists, try reinstalling the package. For further assistance, you can check the Ultralytics documentation. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***&gt;

[glenn-jocher]

Glad to hear you resolved the issue! If you have any more questions or need further assistance, feel free to ask.

[maoxiuyuan]

Thank you for your help! I appreciate your support.

[glenn-jocher]

You're welcome! If you have any more questions or need further assistance, feel free to reach out.

[khanhdc]

First of all, what an amazing job
I also have an question about this version 11. Here you mentioned that

Enhanced Feature Extraction: YOLO11 employs an improved backbone and neck architecture, enhancing feature extraction capabilities for more precise object detection.
May I ask what backbone was deployed in V11, and what make it so powerful?

[glenn-jocher]

@khanhdc thank you for your kind words! YOLO11 uses an enhanced backbone architecture designed for improved feature extraction, contributing to its precision and efficiency. For more details, please visit the YOLO11 documentation.

[xingchen020622]

How to solve it in the latest yolov11 environment TypeError: load_inference_source() got an unexpected keyword argument 'img_size'

[glenn-jocher]

@xingchen020622 the error you're encountering suggests a mismatch in function arguments. Please ensure you're using the latest version of the YOLO11 codebase, as the function signature might have changed. You can check the documentation for updates: YOLO11 Documentation.

[lewismoss]

Hey, I am trying to implement a bird detection system. I would like to feed in images and detect birds. - - Is it possible to set YOLO to only look for birds and ignore everything else?

• Would you recommend some sort of fine-tuning for this purpose?
• Is it possible to detect 'small bird', 'medium bird', 'large bird'?

[vogloblinsky]

Yes you have to annotate a dataset with robotflow if the standard classes of Coco dataset are not enough.
For the size it is a bit more complicated because you need to have a sort of scale of your picture to see if the boxes the model will gives you are small or not.
Or annotate every type of birds you want to detect.

[lewismoss]

Thank you for your reply. Is it possible to set YOLO to only look for birds and ignore everything else?

[seabirdman]

Yes, you can can write a .py function so that Yolo only displays one or more of the 80 'built-in' classes. In fact the example code I attach below goes one step further, not only can you use PickerNamer code to suppress the names of any class of object you can also use this function to rename any class that happens to detect classes you are interested in. Let me explain. In Yolo 8 and now in 11, there is no class for fish. And tho I am an ornithologist I sometimes want to detect fish - but not not so bad as to want to bother with the chore of training. It turns out that Yolo consistently calls fish birds, bananas or kites. So, I wrote PickerNamer (attached as .txt) to let me suppress any classes I want AND to rename any class in my detection script. in your use-case you would just want to use the 'picker' feature and thereby just 'pick' bird class to be the only class to display. But you could also use the renamer feature to change 'bird' to puffin or sparrow or to any other bird name you want. In other words, PickerNamer is a total hack - if you are wanting to have your Yolo actually differentiate bird species - you will have to train a model for that. But, in the interim, see how far PickerNamer can take you. cheers.
PickerNamer.txt tested with yolo8n and yolo11n.

[glenn-jocher]

Yes, you can configure YOLO to detect only birds by filtering the output to include only the bird class. This can be done by specifying the class index for birds in your detection script. For more detailed customization, consider fine-tuning the model on a bird-specific dataset.

[davidetedd9203]

Hi,
i'm trying training yolo11 on a custom dataset. unfortunately dfl_loss and box_loss stays near 1. is it true that a good model should have less than 0.05 for these 2 metrics? if yes what you would suggest to improve training?
thanks in advance

[glenn-jocher]

@davidetedd9203 for improving your YOLO11 model's performance, consider adjusting hyperparameters like learning rate, batch size, or using data augmentation techniques. A good model typically has lower loss values, but achieving less than 0.05 for dfl_loss and box_loss might not always be necessary. Experiment with different configurations to see what best suits your dataset.

[nguyenvandang1998]

Hi!
I have tried yolov11 and it is great, but I want to ask how to train and predict with gray image, and as I understand it will shorten the prediction time, is that right?
Thanks!!

[glenn-jocher]

@nguyenvandang1998 to train and predict with grayscale images using YOLO11, you can convert your images to grayscale before training and adjust the input channels accordingly. While this might reduce the model's complexity slightly, the impact on prediction time is generally minimal.

[nguyenvandang1998]

So what settings do I need to adjust the input channels. I mean what parameters need to be changed to train with grayscale images compared to regular training?

[glenn-jocher]

To train with grayscale images, set the channels parameter to 1 in your dataset configuration file, and ensure your images are preprocessed to grayscale.

[nguyenvandang1998]

Thank you very much!!

[glenn-jocher]

You're welcome! If you have any more questions about using YOLO11, feel free to ask.

[shahedmomenzadeh]

did you published a paper for this model (YOLOv11) ?

[glenn-jocher]

@shahedmomenzadeh ultralytics has not published a formal research paper for YOLO11. For the latest information, please refer to our GitHub repository and documentation.

[davidetedd9203]

hi! how can i know which layer belong to backbone? i want to freeze those layers during training.

[glenn-jocher]

To identify and freeze the backbone layers in YOLO11 during training, you can refer to the model's architecture definition in the configuration files, typically found in the model's YAML file. This will help you locate the specific layers constituting the backbone.

[JackEdoubleS]

I am converting the yolo11n.pt model to .tflite model. After running
"""
model = YOLO("/home/cuonga1/object_detection/runs/detect/train5/weights/best.pt")
model.export(format='tflite')
"""
I got the best.tflite with the outputs format:
name: Type Description Shape Min/ Max

output Unknown (float32) Coordinates of detected objects, [1,34,8400] []/[]
class labels and confident score

But my app need the outputs like:
name: Type Description Shape Min/ Max

location BoundingBox (float32) [] []/[]
category Feature (float32) [] []/[]
score Feature (float32) [] []/[]
Number of detection Feature (float32) [] []/[]

How can i make the designed outputs and can you explain the meaning of [1,34,8400]?

[glenn-jocher]

The number 34 in the output shape corresponds to the 4 box coordinates, 1 confidence score, and 29 class probabilities (assuming indices start from 0). If you have 30 classes, this total becomes 35, as you correctly noted. The number 8400 is related to the number of anchor boxes and grid size, which is dependent on the image size and model architecture, determining how many detections are made. For further insights, reviewing the detection head in the source code at ultralytics/nn/modules/head.py might be beneficial.

[JackEdoubleS]

After running yolo on 30 classes, I export to .TFlite and the model output shape was [1,34,8400]. Since the number was 34 for 30 classes, this is the thing that make me confuse.

[glenn-jocher]

The number 34 in your output corresponds to 4 box coordinates, 1 confidence score, and 29 class probabilities, which seems off for your 30 classes. It should indeed be 35. The 8400 represents the number of potential detections, influenced by the grid size and model architecture, not directly by the number of classes. For precise details, you might want to check the detection head defined in ultralytics/nn/modules/head.py. Adjustments in the model configuration or architecture may be necessary to align with your class count.

[JackEdoubleS]

as I can see in the postprocess function in head.py,
`def postprocess(preds: torch.Tensor, max_det: int, nc: int = 80):
"""
Post-processes YOLO model predictions.

    Args:
        preds (torch.Tensor): Raw predictions with shape (batch_size, num_anchors, 4 + nc) with last dimension
            format [x, y, w, h, class_probs].
        max_det (int): Maximum detections per image.
        nc (int, optional): Number of classes. Default: 80.

    Returns:
        (torch.Tensor): Processed predictions with shape (batch_size, min(max_det, num_anchors), 6) and last
            dimension format [x, y, w, h, max_class_prob, class_index].
    """
    batch_size, anchors, _ = preds.shape  # i.e. shape(16,8400,84)
    boxes, scores = preds.split([4, nc], dim=-1)
    index = scores.amax(dim=-1).topk(min(max_det, anchors))[1].unsqueeze(-1)
    boxes = boxes.gather(dim=1, index=index.repeat(1, 1, 4))
    scores = scores.gather(dim=1, index=index.repeat(1, 1, nc))
    scores, index = scores.flatten(1).topk(min(max_det, anchors))
    i = torch.arange(batch_size)[..., None]  # batch indices
    return torch.cat([boxes[i, index // nc], scores[..., None], (index % nc)[..., None].float()], dim=-1)

since if thenc=80, the tensor shape would be (batch_size, num_anchors, 84)` and there is no confident score here.

Sorry if I have any misunderstandings.

[glenn-jocher]

It seems there might be some confusion about the output tensor dimensions. The shape (batch_size, num_anchors, 4 + nc) should indeed include the confidence score along with class probabilities. If this isn't clear in your setup, please verify how the postprocess function is handling the prediction tensor. The code in head.py is responsible for this processing, and reviewing it may clarify how scores and boxes are extracted and formatted. If further questions arise, feel free to ask!

[abhishek7467]

Hello,
I am trying to find the config.yaml file in yoloV11 where i can change the initial weights accordingly for classes for my dataset.
I am using yoloV11 large segmentation model.

or how i can change initial weights in yoloV11.

[glenn-jocher]

@abhishek7467 to adjust the initial weights for your YOLO11 model, you can modify the weights parameter in your training script or YAML configuration file to point to a custom weights file that aligns with your dataset's classes. If you need further guidance, please refer to the training documentation for detailed instructions.

[ishan-bhattacharya-hyve]

Hi, I would like to use YOLO11 as a feature extractor for similarity measurements. How can I extract the image features from the model? Thanks

[glenn-jocher]

To extract image features using YOLO11, you can modify the model to output intermediate layer features. This typically involves accessing specific layers in the model's architecture. For detailed guidance, please refer to the YOLO11 documentation or explore the model's architecture in the Ultralytics GitHub repository.

[Smartking1]

Is there a way to integrate extended Kalman filters with YOLO11?

[glenn-jocher]

@Smartking1 integrating extended Kalman filters with YOLO11 isn't directly supported, but you can implement it externally by using YOLO11's output for object detection and then applying the Kalman filter for tracking. This approach allows you to enhance the tracking capabilities of YOLO11.

[chuhuiL]

What are segmentation model loss function and the weight?

[glenn-jocher]

@chuhuiL for YOLO11 segmentation models, the loss function typically combines multiple components like classification loss, box regression loss, and mask loss to optimize model performance. The weights for these components can be adjusted based on the task requirements, impacting the model's focus during training. For detailed information on configuring these, you can refer to the YOLO11 Segmentation Documentation.

[GustavoPiaz2102]

How do I access "box_loss" on yolo11?

[glenn-jocher]

@GustavoPiaz2102 to access "box_loss" in YOLO11, you can retrieve it from the training results or logs, typically available during model training. If you need further assistance on how to obtain or use these logs, please refer to the Ultralytics documentation.

[C00276177Tomas]

Hello, does YOLO11 support transfer learning? Freezing the backbone? Can't get it to work

[glenn-jocher]

@C00276177Tomas hello! Yes, YOLO11 supports transfer learning, including freezing the backbone during training. If you're encountering issues, please ensure your setup aligns with the documentation. You can refer to the YOLO11 Train documentation for more detailed instructions. If you need further assistance, feel free to ask!

[C00276177Tomas]

@glenn-jocher Thanks for the speedy reply. I've been continuing to try and get it working but still no luck. Any help for this would be much appreciated.

What I'm trying to achieve is to load the YOLO("yolo11n.pt") model with all 80 classes, then use transfer learning to add new detection classes called helmet and noHelmet to the pretrained model while still keeping the pretrained classes. So in the final model I would have 82 clases. I'm using the following code to train the model:

model = YOLO("yolo11n.pt")
results = model.train(data="/content/MyFolder/data.yaml", epochs=30, imgsz=640, freeze=23)

I've messed around with different values for the freeze varaible and the yaml file generated by roboflow but still no luck. The resulting model always only detects the two classes that Im trying to add and discards the previous classes. If I try to add the pretrained classes by modifying the yaml file it still dosn't work. This is my yaml file from roboflow:

train: ../train/images
val: ../valid/images
test: ../test/images

nc: 2
names: ['helmet', 'noHelmet']

roboflow:
workspace: sitesafeai
project: sitesafemodel3
version: 4
license: CC BY 4.0
url: https://universe.roboflow.com/sitesafeai/sitesafemodel3/dataset/4

[glenn-jocher]

@C00276177Tomas thank you for your detailed query! The issue arises because the data.yaml specifies only 2 classes ('helmet' and 'noHelmet'), which overrides the COCO classes. To preserve the 80 COCO classes and add 2 new ones, you need to modify your data.yaml file to include all 82 classes. Update the nc field to 82 and the names list to include both the COCO classes and your new classes. Additionally, freezing the backbone using freeze should work as expected. Let me know if you continue facing issues!

[C00276177Tomas]

@glenn-jocher Hi again. I've tried both of these examples bellow where I'm adding my classes to the start and end of the list but it still dosn't work. I'm just getting a model that only detects the additional classes.

train: ../train/images
val: ../valid/images
test: ../test/images

nc: 82
names: ['head', 'helmet', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush']

roboflow:
workspace: sitesafeai
project: hard-hat-sample-rh89z
version: 3
license: Public Domain
url: https://universe.roboflow.com/sitesafeai/hard-hat-sample-rh89z/dataset/3

Example 2, helmet and head at the end:

train: ../train/images
val: ../valid/images
test: ../test/images

nc: 82
names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush', 'head', 'helmet']

roboflow:
workspace: sitesafeai
project: hard-hat-sample-rh89z
version: 3
license: Public Domain
url: https://universe.roboflow.com/sitesafeai/hard-hat-sample-rh89z/dataset/3

[glenn-jocher]

To include new classes while retaining the 80 COCO classes in YOLO11, your training dataset must explicitly define the nc and names in the data.yaml to match all 82 classes. Ensure that your annotations include bounding boxes for all 82 classes across your training images (not just the new ones). Without labels for the COCO classes, the model perceives them as absent, effectively forgetting them during training. Consider merging COCO annotations with your custom dataset to resolve this.

[hetfmt]

I have loaded the "YOLO11m-cls.pt" model and ask for the detailed information (see attached image).

How come the layer-listing (left-most column) ends at "148" while the summary claims there are "187" layers?
If I want to freeze layers for transfer-learning, should I use the layer-listing (left, ends at 148) to specify the layers to freeze?

[glenn-jocher]

@hetfmt the discrepancy between the layer listing and summary is due to model architectural details—some additional layers like activation functions or merged components may not be explicitly listed. For freezing layers during transfer learning, refer to the layer indices in the summary (total layers) rather than the visible layer listing. Let me know if you'd like further guidance!

[hetfmt]

Thank you for the fast reply!
So, to freeze everything but the "head", I essentially count backwards to find the range of layers to freeze? Would that be correct?

[glenn-jocher]

Yes, to freeze all layers except the "head," you can start by counting the total layers in the model summary (e.g., 187 for YOLO11m-cls) and exclude the last few layers (corresponding to the head) from freezing. Use the layer indices from the summary as the reference point. Let me know if you need further help with this!

[hetfmt]

Having loaded and printed the "YOLO11m-cls.pt" model, I see that there is a "section" (the head) called "Classify".

I would like to exchange this Classify-head with a regression-head, but I can't figure out how to do it for this model? I have read the above comments, and other similar threads for YOLOv8, but none of this brings me closer to figuring out how to exchange the head of the mentioned classification-model. Can you please advice me how to make that change to the network?

[glenn-jocher]

@hetfmt to replace the classification head in YOLO11 with a regression head, you will need to customize the model architecture. This involves creating a custom head layer with a regression loss function and modifying the final layers of the model. Currently, the YOLO API doesn't natively support automatic replacement of specific heads, so you may need to manually extend or modify the model architecture in the codebase or train it with a custom configuration. For detailed guidance, refer to Ultralytics' source code on GitHub and explore the API and architecture definitions.

[hetfmt]

@glenn-jocher Thank you. I'll see if I can figure it our :-)

[glenn-jocher]

You're welcome! If you encounter any issues or have further questions while making modifications, feel free to reach out. The YOLO community and resources like the Ultralytics documentation are always here to help. Good luck!

[slamivait]

I am trying to create educational content based on a project I'm doing, and have made a diagram of yolov11 and am hoping to get feedback on if it is accurate. I don't see a way to upload it here, but am happy to send it to someone if that's an option. The diagram I have is very similar to this post: https://medium.com/@nikhil-rao-20/yolov11-explained-next-level-object-detection-with-enhanced-speed-and-accuracy-2dbe2d376f71
There are two differences. One difference is I have the SPPF block and the C2PSA in the backbone to be consistent with the Yolov8 diagram having the SPPF in the backbone. (Which is correct, to have SPPF and C2PSA in the neck or the backbone?) The second difference, and what I'm most unsure about is the head. The paper describes "The head of YOLOv11 includes several CBS (Convolution-BatchNorm-Silu) [19] layers after the C3k2 blocks." I'm not seeing where to find this in the code, so I'm not sure what several means. Is this one after each resolution output from a C3k2 into the head? Or is it several after each C3k2 block? The paper also says "Each detection branch ends with a set of Conv2D layers". So, my inclination is at each resolution detection in the head to put a CBS, then a Conv2D, then "Detect" which would indicate where the Bounding Box, Objectiveness Score, and Class Score are determined. Would this be accurate? Should it be described differently? Others are simply stating "Detect" in the head at each resolution level, but it feels like it's missing the CBS and Conv2D described in the paper.

[glenn-jocher]

@slamivait thank you for your detailed question! Regarding the SPPF and C2PSA, in YOLO11, the SPPF block is part of the neck, not the backbone. For the head, the architecture includes a series of CBS layers after each C3k2 block, followed by Conv2D layers for each resolution, ending in the "Detect" layer. Your interpretation aligns well with the described structure, but for precise implementation details, I recommend reviewing the YOLO11 model code in the Ultralytics repository. Let me know if you have further questions!