Day-to-Night Translation using CycleGAN

Write Up (PDF) Code Colab

Method Overview

CycleGAN Day↔Night method overview
CycleGAN with LSGAN adversarial losses, cycle-consistency and identity losses. Trained on BDD100K (unpaired Day/Night split).

Abstract

We study unpaired image-to-image translation for converting daytime driving scenes into nighttime equivalents using CycleGAN. The model learns mappings between Day↔Night without aligned pairs by combining adversarial training with cycle-consistency and identity losses. We report distributions of SSIM/PSNR/LPIPS and their trends across epochs, and show qualitative results for both Day→Night and Night→Day mappings.

Visual Comparisons

Day input Generated Night
Day (Input)
Night (Generated)
Night input Generated Day
Night (Input)
Day (Generated)

Results

Histograms of SSIM, PSNR, LPIPS for A→B and B→A
Histograms across the test set for A→B (Day→Night) and B→A (Night→Day): SSIM, PSNR, LPIPS.
SSIM, PSNR, LPIPS trends across epochs
Score vs. epoch trends showing improving SSIM/PSNR and decreasing LPIPS.