Project · Course project
Road Surface Classification
An image classifier that predicts whether a road is dry, snowy, icy or mixed from a dashcam photo.
- Machine learning
- PyTorch
- Computer vision
- Course project
- Team of three
- Role
- Built my own model, its experiments and the VGG16 comparison
- When
- Mar – Apr 2026
- Team
- Three students
- Data
- 9,000 dashcam images
Overview
A team project for a machine learning course: teach a model to tell dry, snowy, icy and mixed roads apart, using dashcam images from the WinterSim Initiative.
On a team of three, I built my own model from scratch in PyTorch, ran six documented training experiments and compared the results against a pretrained VGG16.
Highlights
Accuracy by run Run Test set Outside images 1 98.3% 23.3% 2 81.5% 25.0% 3 90.3% 21.7% 4 92.3% 16.7% 5 69.0% 26.7% 6 75.2% 33.3% Narrowed the gap
Over six runs, test accuracy fell from 98% to 75% while accuracy on outside images rose from 23% to 33%, as the model memorized less.
Mixed-road accuracy, test set Before 1.6% After weighted loss 70.5% Fixed an ignored class
Weighting the loss lifted accuracy on the rare mixed class.
Accuracy on outside images My model 33.3% Pretrained VGG16 36.7% Found the real limit
A pretrained VGG16 barely beat my small model on outside images, which pointed to the data, not the model, as the limit.
Key challenges
98% accuracy that meant nothing Result: Test accuracy dropped to 82% while outside accuracy held, which confirmed the 98% was memorization. From then on, I judged every run on outside images.
Problem
The first model scored 98% on the test set but only 23% on 60 outside images it had never seen.
What I did
Traced it to the model memorizing near-identical video frames. On a teammate's suggestion, I shrank it from 1.8 million to about 3,000 parameters and retested on outside images.
How it's built
Model
- PyTorch
- torchvision
Training
- Jupyter
- Google Colab
- Kaggle GPUs
More projects
Back to all work
VolleyNet
- Mobile app
- iOS and Android
- Flutter
- Supabase
- Stripe
Social app for volleyball players and hosts, live on iOS and Android with 500+ users. Players find teammates, chat and keep live scores; hosts run paid events with waivers and reminders.
Interactive Idlewild
- Mobile app
- Flutter
- Firebase
- Built solo
- Honors thesis
Mobile app preserving the history of Idlewild, Michigan, a historic Black resort community. Visitors explore an illustrated map, read community-written stories and find events, while admins review every post.