INSAIT, part of Sofia University “St. Kliment Ohridski” introduces EgoNight – the first large-scale study and benchmark dedicated to understanding how modern AI systems perceive and interpret visual information at night. The work has been accepted at ICLR 2026, one of the leading global conferences in artificial intelligence.
As technologies such as smart glasses, autonomous robots, and intelligent assistants become increasingly integrated into everyday life, the ability of AI systems to operate reliably in real-world conditions is critical. However, most existing computer vision models are trained predominantly on daytime data, limiting their effectiveness in low-light environments.
EgoNight addresses this gap through a unified framework for evaluating egocentric vision at night. At its core is EgoNight-VQA, a visual question answering task that combines aligned day–night videos, depth estimation, and cross-illumination retrieval.
The dataset includes 90 videos captured in Sofia, along with synthetic sequences, totaling 3,658 question–answer pairs across 12 categories. The data has been validated through more than 300 hours of human annotation. A key contribution is a novel day–night alignment pipeline, which leverages daytime references to improve annotation quality under low-light conditions.
Experimental results reveal a substantial performance gap: even state-of-the-art multimodal models achieve only around 30% accuracy on EgoNight-VQA. The findings show that the challenge extends beyond poor visibility, highlighting limitations in reasoning and cross-modal understanding in nighttime scenarios.
At the same time, the study demonstrates the promise of synthetic data as a scalable approach to improving AI performance in low-light conditions.
EgoNight will be presented in person at ICLR 2026 in Brazil.
Resources:
Project page: https://insait-institute.github.io/EgoNight/
Paper: https://arxiv.org/abs/2510.06218
Code: https://github.com/insait-institute/EgoNight
Dataset: https://huggingface.co/datasets/INSAIT-Institute/EgoNight/