AI/MLData ScienceResearchGroup Project2026

Pothole Segmentation and Detection

A classical machine learning system for detecting and segmenting potholes in road images, developed as two comparative pipelines within a computer vision course project. No deep learning was used, only handcrafted features and classical models.

Machine Learning Engineer, Pipeline 2 (Boosting): feature engineering, model development, dataset augmentation, and Streamlit deployment

Problem

Manual pothole inspection is slow and inconsistent, and the competition dataset (ARA 7.0) alone was too small to train a reliable classical model, limiting how well the first pipeline could generalize to new road images.

My Contribution

I led Pipeline 2, a gradient boosting classifier built on 36 engineered features, and extended the original ARA 7.0 training data with hard negative examples from the RDD2022 India dataset to improve generalization. This built on Pipeline 1, a Random Forest and SLIC superpixel segmentation approach developed by a teammate as the project baseline. I also handled the Streamlit and Hugging Face deployment for Pipeline 2.

Impact

Augmenting the training data with hard negative examples helped Pipeline 2 generalize more reliably than the original baseline, and both pipelines were deployed as live web apps so the comparison could be explored directly rather than only reported.

Lesson Learned

Careful feature engineering and thoughtful dataset augmentation can meaningfully improve a classical machine learning pipeline, even without moving to deep learning.

Tech Stack

PythonScikit-learnSLIC Superpixel SegmentationRandom ForestGradient BoostingFeature EngineeringStreamlitMachine Learning

GitHub (Pipeline 1: RF + SLIC)GitHub (Pipeline 2: Boosting)Live Demo (Pipeline 1)Live Demo (Pipeline 2)Hugging Face Demo (Pipeline 2)Report Demo Video

Demo Video

Project Media

Pothole Segmentation and Detection poster

Back to all projects