A survey of day-night illumination domain translation for outdoor vision: Methods, datasets, and evaluation protocols

Alam, Md Shadab; Singh, Priyanshu; Bazilinskyy, Pavlo · 2026 · Crossref

DOI: 10.1007/s00138-026-01843-8

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This survey addresses the performance degradation of computer vision systems in outdoor environments due to day-night illumination shifts. Low illumination, mixed lighting, glare, and sensor noise significantly weaken cues for detection, segmentation, and tracking, creating a domain gap between daytime training data and nighttime deployment conditions. The authors focus on Illumination Domain Translation (IDT), which maps images between day and night domains while preserving geometry, semantics, and temporal coherence. The study aims to systematize methods, datasets, and evaluation protocols, distinguishing IDT from general low-light enhancement or domain adaptation by emphasizing the preservation of scene content under severe illumination ambiguity. The authors developed a constraint-centric taxonomy to organize the literature, linking supervision types, five domain gap factors (illumination, glare, noise, weather, motion), and five families of constraints (cycle/identity, semantic, physics-informed, correspondence, temporal) to specific failure modes. They reviewed 30 representative methods and 23 datasets, categorizing them based on their primary modeling commitments. Additionally, the study conducted an artifact availability audit of 34 published methods to assess reproducibility. The authors propose a four-part evaluation protocol (P/S/D/T) covering perceptual quality, structural fidelity, downstream task utility, and temporal stability, arguing that current evaluations are overly biased toward visual realism. Key findings reveal that unpaired training dominates the field (23 of 30 methods), necessitating strong constraints to ensure semantic faithfulness. While cycle and identity constraints are common, explicit semantic or temporal constraints are less frequent, with only three methods addressing temporal stability. Evaluation evidence is heavily skewed toward perceptual realism (22 of 30 papers), with limited reporting on semantic preservation (7 of 30) or downstream utility (9 of 30). Furthermore, while most methods address general illumination shifts, specific nighttime challenges like glare, noise, and motion are under-covered. The artifact audit showed that while 29 of 34 methods release code, only 21 specify licenses and 19 provide full reproducibility packages, hindering practical deployment. The significance of this work lies in its structured framework for understanding IDT as a distinct problem requiring explicit constraints to prevent semantic drift and hallucination. By mapping gap factors to algorithmic solutions, the survey provides actionable guidance for selecting methods based on specific failure modes. The proposed P/S/D/T evaluation protocol highlights critical gaps in current research, particularly the lack of evidence regarding downstream utility and temporal stability. The authors conclude that progress in IDT depends on rigorous evaluation of semantic preservation and reproducibility, rather than relying solely on perceptual metrics, to ensure robust performance in safety-critical outdoor vision applications.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• dark adaptation mesopic