Trained a Custom YoloV11 OBB Model with Custom Dataset, But Getting No Detections

[BeaverBytes]

Environment

python=3.12.7
torch=2.5.1
torchvision=0.20.1
ultralytics=8.3.38
ultralytics-thop-2.0.12
opencv-python=4.10.0.84
nvidia-cuda-cupti-cu12=12.4.127
nvidia-cuda-nvrtc-cu12=12.4.127
nvidia-cuda-runtime-cu12=12.4.127

Dataset

Object detection dataset was created using cvat. The dataset consists of Keurig images and the labeled objects are the buttons on the coffee machine. I modified the data.yaml to fit the YoloV11 data formatting. The dataset has the following structure:

• images/
  • train/*.png
  • val/*.png
• labels
  • train/*.txt
  • val/*.txt
• data.yaml

data.yaml has the following contents

Train: Train.txt
names:
  0: 4oz
  1: 6oz
  2: 8oz
  3: 10oz
  4: 12oz
  5: strong
  6: iced
  7: hot
  8: add_water
  9: descale
  10: handle
  11: water bin
  12: power
path: .
train: images/train
val: images/val

Example of what the labeled contents will have:

4 0.590707 0.701204 0.687732 0.701204 0.687732 0.756579 0.590707 0.756579
9 0.524006 0.534504 0.633152 0.534504 0.633152 0.575587 0.524006 0.575587
8 0.365464 0.542247 0.508409 0.542247 0.508409 0.588087 0.365464 0.588087
5 0.359874 0.598838 0.511762 0.586760 0.519713 0.633986 0.367825 0.646064
7 0.509094 0.587168 0.660982 0.575090 0.668933 0.622316 0.517045 0.634393
0 0.426112 0.659534 0.511870 0.659534 0.511870 0.704180 0.426112 0.704180
1 0.537865 0.648219 0.611502 0.648219 0.611502 0.698229 0.537865 0.698229
2 0.385390 0.716093 0.473756 0.716093 0.473756 0.772652 0.385390 0.772652
3 0.486745 0.709545 0.576848 0.709545 0.576848 0.763725 0.486745 0.763725
6 0.642680 0.636316 0.712857 0.636316 0.712857 0.685132 0.642680 0.685132
12 0.428351 0.364767 0.510677 0.364767 0.510677 0.417925 0.428351 0.417925

Code

This is the code I used to train my custom YoloV11 OBB model with my custom dataset

import cv2
from ultralytics import YOLO
model = YOLO("yolo11n.yaml")
model.train(data="<path_to_datasets>/datasets/v19/data.yaml", epochs=10, imgsz=640)

model = YOLO("/home/ncbernar/.pyenv/runs/detect/train5/weights/best.pt")
results = model("<path_to_keurig_image>/keurig.png")

for r in results:
    im_array = r.plot()  # plot a BGR numpy array of predictions
    im = cv2.cvtColor(im_array, cv2.COLOR_BGR2RGB)  # convert to RGB for display
    cv2.imshow("Detection Results", im)
    cv2.waitKey(0)  # wait for a key press to close the window

cv2.destroyAllWindows()

After results is collected, I get the following message:

image 1/1 <path_to_image>/keurig_top_view-Copy1.png: 640x448 (no detections), 34.7ms
Speed: 1.6ms preprocess, 34.7ms inference, 0.4ms postprocess per image at shape (1, 3, 640, 448)

To reproduce this issue, please use the following dataset from here.

Problem

As shown above, results tells me that the model was unable to detect any of the objects, even though the labels and images were provided as part of the dataset. One example of what the labeled image looks like is below:

But after I feed the model an image for inference, I get the following result with no bounding boxes:

Discussion

Basically, I expected the model to be able to detect the obejcts, or buttons, in the image. It acknowledged the dataset and labels, but still no sign of the bounding boxes being shown on the resulting image.

Any reason for why this happens? I suspect that my dataset may not be correctly formatted for the OBB model, but it could be a different issue.

[pderrenger]

👋 Hello @BeaverBytes, thank you for sharing details about your custom training experience with Ultralytics 🚀! We understand the effort you’ve put into setting up your custom YoloV11 Object Bounding Box (OBB) model, and we’re here to help.

While we review your discussion, here are a few suggestions that might help you identify the issue:

If this is a 🐛 Bug Report, please make sure to provide a minimum reproducible example (MRE), including a representative data sample and code snippet to allow us to replicate the issue on our end. If your provided example includes all necessary information, thank you!

If this is a custom training ❓ Question, it's important to ensure your dataset adheres to YOLO OBB label formatting and that you're following our Tips for Best Training Results. Since OBBs require a specific annotation format, double-check that all bounding box corners are correctly defined in your label files.

Docs and Community Resources

• 📚 For training and inference guidance, visit the Ultralytics Docs.
• For OBB-specific tutorials, confirm you’ve reviewed examples pertinent to your model configuration under the YOLO modes.
• Engage with our vibrant community to exchange ideas and solutions:
  • Discord 🎧 for real-time discussions.
  • Discourse for detailed questions and solutions.
  • Subreddit for shared insights and advice.

Verify Installation and Dependencies

Please ensure you're running the latest version of ultralytics along with updated dependencies required for optimal performance on systems like CUDA, PyTorch>=1.8, and Python>=3.8. Update your Ultralytics installation with the following:

pip install -U ultralytics

Additionally, confirm that your system and Python packages match the required versions specified in the requirements file.

Environments

Test your training and inference setup on validated environments to ensure consistency:

• Free cloud GPUs:
• Cloud services like GCP (Quickstart) and AWS (Quickstart).
• Docker Images for isolated environments:

Training Suggestions and Debugging Checklist

• Verify your data.yaml file is properly formatted and paths are correct.
• Ensure your label files strictly use YOLO OBB standards for annotation.
• Review training logs for any anomalies during training that could indicate a potential misconfiguration, overfitting, or underfitting.

Status

Our Continuous Integration (CI) tests verify the latest commit’s functionality on macOS, Windows, and Ubuntu. Check the project's current CI status:

---

This is an automated response to get you started 😊. An Ultralytics engineer will review your question and provide additional help soon!

[Y-T-G]

1. yolo11n.yaml trains a normal detection model, not OBB. And start training from "yolo11n-obb.pt" not yaml. https://docs.ultralytics.com/tasks/obb/
2. 10 epochs is not enough. Set it to 100.
3. You probably don't have enough images either. Follow the guidelines here: https://docs.ultralytics.com/yolov5/tutorials/tips_for_best_training_results/

[BeaverBytes]

Ahh sorry, I meant to put "yolo11n-obb.yaml", not "yolo11n.yaml". But I will give this a try! Thanks for the response :)

[BeaverBytes]

Sir, you are a g. I'm able to see the labels on my images now thanks to your advice. Basically, I just transferred the weights from the pretrained model to the yaml file where I want to build a custom model based on my dataset and trained the model with 100 epochs.

Thanks again for the help!

[glenn-jocher]

You're very welcome! Great to hear it's working now. All credit goes to the YOLO community and the Ultralytics team for making these tools so effective. Happy detecting!