This project uses a YOLOv11 object detection model to identify bike lane infrastructure and related objects in urban street images.
YOLOv11 was chosen for its balance of speed and accuracy in real-time object detection tasks.
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 model rather than trained from scratch, which allows it to perform reasonably well even with a small dataset.
The goal of this project was not only to train a model, but to understand how dataset quality and structure affect performance in real-world computer vision tasks.
Intended Use Cases:
Exploring bike lane infrastructure in street imagery
Supporting transportation or urban planning research
Analyzing cyclist environments and road conditions
The dataset contains 147 images of urban street environments with varying road layouts, lighting conditions, and traffic scenarios.
Class Distribution
Class
Count
Vehicle
253
Bicycle Lane
129
Shared Dotted Lane
124
Solid Lane
59
Cyclist
13
Bicycle
2
Car
2
This dataset shows strong class imbalance, where some classes appear very frequently while others have very few examples. This directly affects model performance.
Annotation Process
The dataset included pre-existing YOLO-format bounding box annotations.
Although the dataset was pre-annotated, I reviewed samples in order to check for consistency and quality. I observed that some classes such as "car" and "vehicle" overlap conceptually, which may introduce ambiguity during training. No major corrections were made, but this overlap influenced how results were interpreted.
I reviewed a subset of images to validate annotation quality, focusing on:
alignment of bounding boxes
consistency of class labels
No major corrections were made. This allowed me to focus on model training and evaluation, but it also represents a limitation since annotation quality was not significantly improved.
This project therefore, emphasizes evaluation and understanding of model performance rather than dataset refinement which is something I wanted in this process.
Dataset Split
Train: 102 images (69%)
Validation: 20 images (14%)
Test: 16 images (11%)
Data Augmentation
Default YOLO augmentations were applied during training:
horizontal flipping
color adjustments
mosaic augmentation
Known Dataset Limitations
Strong class imbalance
Extremely small sample sizes for some classes
Limited total dataset size
Mostly daytime, urban conditions
Training Procedure
The model was trained using the Ultralytics YOLOv11 framework in Google Colab.
Training was performed in Google Colab using an RTX 3070. Training took approximately 5-10 minutes.
Training Details:
Framework: YOLOv11 (Ultralytics)
Epochs: 50
Batch size: 16
Image size: 640 × 640
Environment: Google Colab
Evaluation Results
Key Metrics
Precision: ~0.88
Recall: ~0.38
mAP50: ~0.48
According to the results, these metrics show that the model is highly precise but has low recall.
This means:
The model is usually correct when it makes predictions
It misses many objects especially harder or less frequent ones
Common classes such as "Vehicle" achieved higher precision and recall, while underrepresented classes like "Bicycle" and "Car" performed poorly due to limited training samples.
This made it so that the performance differences across classes were influenced by class imbalance, with larger classes performing more reliably.
Performance varied across classes, as shown in the table below:
Class
Relative Performance
Vehicle
High
Bicycle Lane
Moderate
Shared Lane
Moderate
Cyclist
Low
Bicycle
Very Low
Example Predictions
Prediction
This example shows the model’s predictions on a validation images. In my analysis I found that the model correctly identifies several road features, but also makes some errors, such as detecting a building as a vehicle and labeling a bus lane as a bike lane. This reflects the model’s tendency to rely on visual similarity and highlights limitations in distinguishing similar structures.
Confusion Matrix
Confusion Matrix
The confusion matrix highlights where the model struggles, particularly between similar lane types and rare classes.
Training Results
Training Results
The training curve shows steady learning, but performance plateaus due to the dataset limitations.
Failure Example
Failure Example
This image shows two types of errors made by the model.
In the first image, the model incorrectly detects a vehicle where there is actually part of a building. This is likely because the building has visual features such as rectangular shapes and edges that resemble vehicles in the training data.
The second model identifies a bike lane where the road marking appears to be a bus lane. This suggests that the model has difficulty distinguishing between different types of lane markings, especially when they share similar visual patterns.
These errors highlight an important limitation, the model relies heavily on visual similarity rather than deeper contextual understanding. Since the dataset contains limited variation and strong class imbalance, the model may generalize incorrectly when encountering unfamiliar scenes.
Evaluation
The model was evaluated using standard object detection metrics.
The precision and recall curves illustrate the model’s ability to detect bike lanes while minimizing false positives. The F1 curve shows the optimal balance between precision and recall, and the PR curve summarizes overall detection performance across confidence thresholds.
Curves
Precision Curve
Recall Curve
PR Curve
F1 Curve
Performance Analysis
The model performs best when:
lane markings are clearly visible
lighting conditions are consistent
objects are large and unobstructed
The model struggles when:
markings are faded or unclear
objects overlap or are partially blocked
objects are small or rare in the dataset
This suggests that dataset quality and balance are more important than model complexity in this case.
Limitations and Biases
Failure Cases
Missed detections of cyclists and small objects
Confusion between similar lane types
Reduced accuracy in cluttered scenes
Data Biases
Overrepresentation of vehicles
Underrepresentation of bicycles and cars
Limited environmental diversity
Environmental Limitations
The model may perform poorly under:
low lighting
occlusion
worn or faded lane markings
Additional Observations
The model sometimes misclassifies lane types such as solid vs shared lane, when markings are partially broken or unclear. This suggests the model relies heavily on strong visual patterns.
Inappropriate Use Cases
This model should not be used for:
autonomous driving systems
real-time safety decisions
high-risk environments
Sample Size Limitations
Some classes like bicycles and car have extremely limited training data, making reliable detection difficult. This contributes directly to low recall.
Final Reflection
I found that this project demonstrates that model performance is heavily dependent on dataset quality.
Even with a strong model like YOLOv11, issues such as:
class imbalance
small dataset size
annotation limitations
It can significantly impact results.
In the end this project to me highlights the importance of data quality, not just model choice, in computer vision applications.
