Update README.md

README.md

@@ -17,30 +17,37 @@ Labels: Binary classification — wildfire present or not.
 Images: Captured under real-world conditions, including diverse environments and challenging scenarios like smoke, clouds, and varying lighting.  
 Size: ~33 000 image labeled images, well-suited for training and validation of computer vision models. 
 - 28,103 images with smoke
-- 31,975 smoke instances
+- 31,975 smoke instances  
 This dataset is formatted to be compatible with the Ultralytics YOLO framework, enabling efficient training of object detection models.   
 Usage: Ideal for fine-tuning state-of-the-art models for wildfire detection tasks.  
 
 ## Model Development Plan
 Model Choice: YOLOv11s  
-Why YOLOv11s?
+Why YOLOv11s?  
 Efficiency: YOLO (You Only Look Once) models are known for their high-speed performance and accuracy, ideal for real-time applications.  
 Versatility: YOLOv11s builds upon prior versions, improving object detection, handling small objects, and performing well under challenging visual conditions.  
 Frugality: Optimized for computational efficiency, aligning with the sustainability goals of the Frugal AI Challenge.
  
 ## Data Preprocessing:
-
-
 Fine-Tuning YOLOv11:
-Load a pre-trained YOLOv11s model as a starting point (transfer learning).
-Replace the output layer to align with the binary classification task (wildfire vs. no wildfire).
+Load a pre-trained YOLOv11s model as a starting point (transfer learning).  
+Replace the output layer to align with the binary classification task (wildfire vs. no wildfire).  
 Train the model using the PyroNear/pyro-sdis dataset.
 
 ## Evaluation:
-
-Metrics:
-Precision and recall to assess detection accuracy.
+Metrics:  
+Precision and recall to assess detection accuracy.  
 Inference time to evaluate real-time feasibility.
 
 Using CodeCarbon, the model's carbon footprint and energy consumption will be tracked, this information will help ensure the model's alignment with the sustainability objectives of the Frugal AI Challenge.
 
+## Results:
+Since evaluation of this challenge is on 20% of the train dataset, we need to be vigilant when talking about model performance.  
+Indeed, a high accuracy in the part of the train dataset could hide the over-fitting of the model.
+
+That is what I've been dealing with when reaching 0.91 accuracy of the model and 0.81 mean_iou on the train dataset: 
+I can see over-fitting while over-performance on the train dataset (see picture below)  - model is used for submission and is called yolo11_tr_20012025_frugalAIchal.pt
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/666354284044e2b1c3287c22/LptPoEeSGH22MG_XftxdP.png)
+
+As a data scientist, another model has been trained and evaluated on the whole "val" dataset.
+