Mitigating Illumination-Induced Domain Shift in Night-Time Pedestrian Detection for Intelligent Vehicles using Annotation-Preserving Diffusion Augmentation

Name: Mitigating Illumination-Induced Domain Shift in Night-Time Pedestrian Detection for Intelligent Vehicles using Annotation-Preserving Diffusion Augmentation
Uploaded: 2026-10-07T23:00:32.557875+00:00
Duration: 674 s
Description: Contrastive-SDXL is a framework that translates daytime images into realistic night-time images while keeping pedestrian details intact. It uses SDXL and fine-tuning with LoRA, guided by DINOv2 features and a detector loss, to create synthetic night data. This significantly improves the performance of pedestrian detectors trained on these augmented images.

Contrastive-SDXL is a framework that translates daytime images into realistic night-time images while keeping pedestrian details intact. It uses SDXL and fine-tuning with LoRA, guided by DINOv2 features and a detector loss, to create synthetic night data. This significantly improves the performance of pedestrian detectors trained on these augmented images.

Open the full episode