Update README.md

README.md

@@ -1,28 +1,29 @@
 Model Description
 
-This model detects bike lane infrastructure and related objects in street images using object detection.
+This project uses a YOLOv11 object detection model to identify bike lane infrastructure and related objects in street images.
 
-I used a YOLOv11 model, which predicts bounding boxes and class labels for objects such as bike lanes, lane markings, cyclists, and vehicles. The model was fine-tuned from a pre-trained version using transfer learning.
+The model detects features such as bike lane markings, shared lanes, cyclists, and vehicles using bounding boxes and class labels. It was fine-tuned from a pre-trained YOLO model rather than trained from scratch, which allows it to learn from a relatively small dataset.
 
-The goal of this project was to understand how well object detection performs in this setting and to evaluate its limitations, rather than just achieving high performance.
+The main goal of this project was not just to build a high-performing model, but to understand how well object detection works in this context and what limitations arise when working with real-world, imperfect data.
 
 Intended Use Cases:
 
-Transportation research
+Exploring bike lane infrastructure in street imagery
 
-Bike lane detection from street images
+Supporting transportation research
 
-Infrastructure analysis
+Analyzing road design and cyclist environments
+
+This model is best suited for exploratory or research purposes rather than real-world deployment.
 
 Training Data
 
 Dataset Source:
 Roboflow Universe – Bike Lane Computer Vision Dataset
 
-Dataset Size:
-147 images, 7 classes
+The dataset consists of 147 images of urban street environments, including a mix of road layouts, traffic conditions, and lighting scenarios.
 
-Class Distribution:
+Classes and Distribution:
 
 Class	Count
 Vehicle	253
@@ -33,21 +34,26 @@ Cyclist	13
 Bicycle	2
 Car	2
 
-Data Collection Methodology:
-The dataset consists of urban street images containing bike lanes, vehicles, and cyclists under various lighting and road conditions.
+One of the most important characteristics of this dataset is strong class imbalance. Some classes, like vehicles and lane markings, appear frequently, while others like bicycles and cars have almost no examples. This has a direct impact on model performance.
+
+Data Collection & Characteristics:
+Images represent real-world urban roads, primarily in daytime conditions, with varying visibility of lane markings and objects.
+
+Annotation Process
+
+The dataset included pre-existing YOLO-format bounding box annotations.
 
-Annotation Process:
-The dataset included pre-existing YOLO bounding box annotations. These annotations label objects using rectangular bounding boxes and class labels.
+Instead of creating new annotations, I focused on reviewing and validating the existing ones. I manually inspected a subset of images to check:
 
-I reviewed a subset of images to verify:
+whether bounding boxes aligned correctly with objects
 
-bounding box alignment
+whether labels were applied consistently
 
-label consistency
+No major corrections were made. While this allowed me to focus on model training and evaluation, it also represents a limitation, since annotation quality was not improved or standardized further.
 
-No major modifications were made to the annotations. This means that while the dataset was usable, the project relied on existing annotations rather than adding new ones.
+This is important because errors or inconsistencies in annotations can directly affect model performance, especially for less frequent classes.
 
-Dataset Split:
+Dataset Split
 
 Train: 102 images (69%)
 
@@ -55,46 +61,47 @@ Validation: 20 images (14%)
 
 Test: 16 images (11%)
 
-Data Augmentation:
-Default YOLO augmentation was used, including flipping and color variation.
+Data Augmentation
 
-Dataset Limitations / Biases:
+Default YOLO augmentation techniques were used during training, including:
 
-strong class imbalance
+horizontal flipping
 
-very limited examples for some classes
+color variation
 
-mostly urban, daytime conditions
+mosaic augmentation
 
-Training Procedure
+Known Dataset Limitations
+
+Significant class imbalance
+
+Extremely small number of examples for some classes
 
-Framework:
-Ultralytics YOLOv11
+Limited dataset size overall
 
-Training Approach:
-Fine-tuning a pre-trained model
+Mostly urban, daytime conditions (lack of environmental diversity)
 
-Hardware:
-Google Colab (CPU or GPU — update if needed)
+Training Procedure
+
+The model was trained using the Ultralytics YOLOv11 framework in Google Colab.
 
-Training Time:
-Approximately ~1 hour
+I fine-tuned a pre-trained model for 50 epochs using images resized to 640 × 640 pixels.
 
-Hyperparameters:
+Training Details:
+
+Framework: Ultralytics YOLOv11
 
 Epochs: 50
 
 Image size: 640
 
-Batch size: 16 (default)
-
-Learning rate: default YOLO setting
+Batch size: 16
 
-Preprocessing:
+Learning rate: default YOLO settings
 
-images resized to 640×640
+Environment: Google Colab
 
-normalization handled automatically
+Training relied on transfer learning, which is especially useful given the small dataset size.
 
 Evaluation Results
 
@@ -106,80 +113,102 @@ Recall: ~0.38
 
 mAP50: ~0.48
 
-Rather than focusing only on the numbers, these metrics help explain how the model behaves.
+Rather than focusing only on these numbers, it is more important to understand what they reveal about the model.
 
-The model has relatively high precision, meaning most detections are correct, but lower recall, meaning it misses some objects.
+The relatively high precision indicates that when the model makes a prediction, it is usually correct. However, the low recall suggests that the model is missing a significant number of objects.
 
-Per-Class Performance
-
-Strong performance on common classes such as vehicles and lane markings
+This imbalance between precision and recall shows that the model is somewhat conservative — it avoids false positives but fails to detect more difficult or less frequent objects.
 
-Weak performance on rare classes such as bicycle and car
+Per-Class Performance
 
-This is largely due to the imbalance in the dataset.
+Strong performance on common classes (vehicles, lane markings)
 
-Visual Examples of Classes
+Weak performance on rare classes (bicycle, car)
 
-(Upload images showing each class)
+This is largely due to the extreme imbalance in the dataset.
 
 Key Visualizations
 
-![Prediction](./val_batch0_pred.jpg)
-![Confusion Matrix](confusion_matrix.png)
-
+![Confusion Matrix](./confusion_matrix.png)
+![Training Results](./results.png)
+![Prediction Example](./val_batch0_pred.jpg)
 
 Performance Analysis
 
-The model performs well when:
+The model performs best when:
+
+lane markings are clearly visible
 
-lane markings are clear
+lighting conditions are consistent
 
-lighting is consistent
+objects are not occluded
 
-The model struggles when:
+However, the model struggles in several situations:
 
-lane markings are faded or unclear
+faded or worn bike lane markings
 
-objects are partially occluded
+overlapping or partially blocked objects
 
-classes have very few examples
+rare classes with very limited training data
 
-This demonstrates that model performance is strongly influenced by dataset quality and class balance.
+These results highlight that performance is not just about the model architecture, but heavily influenced by the dataset.
+
+In particular, the lack of examples for certain classes makes it difficult for the model to learn meaningful patterns.
 
 Limitations and Biases
 
-Failure Cases:
+This model has several important limitations that should be clearly acknowledged.
+
+Failure Cases
 
 missed detections of bicycles and cars
 
-errors when lane markings are unclear
+incorrect detections when lane markings are unclear
 
 confusion between similar lane types
 
-Data Biases:
+Data Biases
 
 overrepresentation of vehicles
 
 underrepresentation of rare classes
 
-limited environmental diversity
+limited diversity in environment and conditions
 
-Environmental Limitations:
+Environmental Limitations
 
-poor lighting
+The model may perform poorly under:
+
+low lighting conditions
 
 occlusion
 
-worn lane markings
+faded or damaged road markings
+
+Inappropriate Use Cases
 
-Inappropriate Use Cases:
 This model should not be used for:
 
-real-time safety decisions
+real-time safety systems
 
 autonomous driving
 
-high-stakes applications
+decision-making in high-risk environments
+
+Sample Size Limitations
+
+Some classes (such as bicycle and car) have extremely limited training data, making reliable detection difficult. This directly impacts recall and overall model performance.
+
+Final Reflection
+
+This project demonstrates that even with a strong model like YOLOv11, performance is highly dependent on the dataset.
+
+Rather than focusing only on improving accuracy, this project highlights the importance of:
+
+dataset quality
+
+class balance
+
+annotation reliability
 
-Sample Size Limitations:
-Classes like bicycle and car have too few examples to be reliably detected.
+Understanding these limitations is essential when applying computer vision models to real-world problems.