LUMINA-26: Low-Light Understanding for Modeling and Interpreting Night-time Actions

2026-06-22 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, medium

Summary

Researchers introduce LUMINA-26, a new dataset designed to address challenges in low-light human action recognition, which suffers from poor illumination, noise, and motion ambiguity. This dataset comprises 6,784 video clips across 26 distinct action classes, featuring recordings from 22 subjects in 20 diverse indoor and outdoor environments under natural low-light conditions. Alongside LUMINA-26, they propose Illumi-Net, an Illumination-Adaptive Mixture-of-Experts Network. Illumi-Net utilizes video-level illumination cues to guide adaptive enhancement and transformer-based spatio-temporal feature extraction, integrating expert-conditioned decision fusion. The proposed method achieves superior performance, surpassing prior benchmark results on ELLAR with a Top-1 accuracy of 55.13% and Top-5 accuracy of 78.87%. Furthermore, Illumi-Net establishes a strong baseline on the new LUMINA-26 dataset, achieving a Top-1 accuracy of 75.95% and Top-5 accuracy of 93.58%, providing a practical benchmark for future research.

Key takeaway

For computer vision engineers developing robust action recognition systems, you should consider LUMINA-26 as a new, diverse benchmark for low-light scenarios. Implementing illumination-adaptive techniques, like those in Illumi-Net, can significantly improve model performance in challenging conditions. This approach helps overcome issues like noise and motion ambiguity, ensuring your models generalize better to real-world night-time applications.

Key insights

A new dataset and adaptive network significantly advance low-light human action recognition by addressing data and model limitations.

Principles

• Low-light action recognition requires diverse, realistic data.
• Illumination cues can guide adaptive model enhancement.
• Mixture-of-Experts improves decision fusion.

Method

Illumi-Net uses video-level illumination cues for adaptive enhancement, followed by transformer-based spatio-temporal feature extraction. Expert-conditioned decision fusion then combines outputs for robust low-light human action recognition.

In practice

• Utilize LUMINA-26 as a benchmark dataset.
• Integrate illumination-adaptive enhancement.

Topics

• Low-Light Vision
• Human Action Recognition
• Video Datasets
• Illumi-Net
• Mixture-of-Experts
• Transformer Networks
• Spatio-Temporal Features

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Can Local Vision-Language Models improve Activity Recognition over Vision Transformers? -- Case Study on Newborn Resuscitation — Fine-tuned local VLMs with LoRA significantly outperform Vision Transformers for newborn resuscitation activity recognition.
• AIGS-Net: Compact Illumination Field Modeling via 2D Gaussian Splatting for Fast Low-Light Image Enhancement — AIGS-Net uses adaptive 2D Gaussian Splatting and zero-parameter encoding for ultra-lightweight, fast low-light image enhancement.
• OR-Action: Multi-Role Video Understanding with Fine-Grained Actions — Fine-grained OR action understanding requires explicit temporal modeling and multi-view feature alignment beyond scene graphs.
• HOI-aware Adaptive Network for Weakly-supervised Action Segmentation — Leveraging human-object interaction (HOI) context dynamically improves weakly-supervised action segmentation accuracy for ambiguous actions.
• 🍓Surface Light Tokenizer🍓 👉Apple unveils LITO a novel latent flow matching model enables... — LITO uses latent flow matching to convert images into high-quality 3D models via compact surface light field representations.
• Multi-Modal Hyper-Graph Fusion for Low-Light Crowd Counting — Multi-modal hyper-graph fusion significantly improves low-light crowd counting by integrating RGB, depth, and edge cues.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.