ML Use Cases at Think Big Analytics with Mo Patel & Laura Frølich - #54

Detection Methods

Mo Patel and Laura Frølich explore object detection techniques in traffic scenarios, focusing on car detection. They highlight the challenges of detecting various objects due to data imbalance, which led them to concentrate on cars. Mo explains the use of single shot methods like YOLO and Single Shot MultiBox Detector, which involve predefined boxes to predict object locations and classes 1.

We ended up focusing more on just the car part of it and not so much on detecting all the other classes that we were initially interested in.

--- Mo Patel

The project aimed to enhance assisted driving systems by detecting stopped cars on roads, providing alerts to other vehicles to prevent accidents 2.

   

Training Challenges

Training models for object detection presents significant challenges, particularly with data annotation and class imbalance. Laura Frølich discusses the complexities of annotating data, which involves labeling objects with bounding boxes and segmentation, a task that can be resource-intensive 3.

We actually built a segmentation tool that allowed people to go ahead and draw boxes around cars and pedestrians.

--- Laura Frølich

Additionally, deploying models in production adds another layer of difficulty, requiring considerations of model compression and infrastructure to handle large-scale inference 4.

   

Augmentation Techniques

Data augmentation is crucial for improving model performance and generalization. Mo Patel and Laura Frølich5. By adjusting parameters like brightness, they enhanced the model's ability to detect roads in diverse conditions without additional data 6.

We retrained the model with data augmentation, it was actually able to detect a lot more road.

--- Mo Patel

This technique is supported by best practices codified in libraries like Keras, making it accessible for developers to implement effectively.