CASE STUDY 07
- Academic
- Computer Vision
- Machine Learning
Bringing day and night closer
Domain adaptation with a bidirectional CycleGAN.
Personal / academic project
The project, step by step
−74%accuracy gap between domains
- 01
Day and night images
Caltech Camera Traps
- 02
Adapt the domains
CycleGAN: day ↔ night transformations
- 03
Evaluate the gap
Compare accuracy across both domains
−74% is the reduction in the accuracy gap between domains, not the final accuracy.
Context
Computer Vision project using the Caltech Camera Traps dataset, focused on domain adaptation between daytime and night-time images.
Problem
Daytime and night-time images have different visual conditions. The project aims to reduce the performance gap between these domains.
Data
Images from the Caltech Camera Traps dataset. No before-and-after examples were provided for this presentation.
Approach
I developed a bidirectional CycleGAN for adaptation between day and night domains.
Method
The general principle of a CycleGAN is to learn transformations in both directions with a cycle-consistency constraint. The details of the configuration used remain to be documented.
Results
A 74% reduction in the accuracy gap. This figure describes the reduction in the gap between domains, not the model's final accuracy.
Technologies
PyTorch, OpenCV and CycleGAN.
Lessons and limitations
The quality of transformed images, failure cases and detailed evaluation protocol remain to be documented.