DAR-Net: A Semantically-Aware Transformer Framework for Underwater Diver Activity Recognition

Forum topic · 小凯 · 2026-06-12

Summary

Researchers Sadman Sakib Enan and Junaed Sattar introduce DAR-Net, a transformer-based framework for classifying diver activities in complex underwater scenes, enabling autonomous underwater vehicles (AUVs) to act as collaborative teammates in high-risk environments. DAR-Net couples transformer-based temporal reasoning with pixel-level scene supervision through a semantically guided, multi-loss training strategy that explicitly aligns global activity recognition with local human-robot interaction semantics—a capability especially critical in low-visibility underwater conditions. To address the severe data scarcity in this field, the authors present the first Underwater Diver Activity (UDA) dataset, a foundational resource containing over 2,600 images annotated with pixel-level masks. Controlled experiments demonstrate that DAR-Net achieves promising accuracy in recognizing six distinct diver activities, outperforming state-of-the-art models. The paper, available as arXiv 2606.12374, serves as a pioneering step toward smarter, collaborative underwater robotic systems and establishes key baselines for future research in underwater human-robot interaction.

Overview

Field: Computer Vision
Authors: Sadman Sakib Enan, Junaed Sattar
Published: 2026-06-10
arXiv: 2606.12374

Abstract

Effective multi-human-robot collaboration is essential for expanding human-led operations in the challenging and high-risk underwater environment. For autonomous underwater vehicles (AUVs) to become true teammates, they must be able to comprehend their surroundings and recognize a diver's activities to offer assistance and ensure safety. Towards this goal, the authors introduce DAR-Net, a novel transformer-based framework that analyzes complex underwater scenes to classify diver activities.

Key Contributions

Semantically guided learning formulation: DAR-Net couples transformer-based temporal reasoning with pixel-level scene supervision.
Multi-loss training strategy: It explicitly aligns global activity recognition with local human-robot interaction semantics, which is particularly critical in low-visibility underwater conditions.
UDA dataset: To address significant data scarcity in this domain, the paper presents the first Underwater Diver Activity (UDA) dataset—a foundational resource containing over 2,600 images with pixel-level mask annotations.

Results

Through rigorous experimental evaluation in controlled environments, DAR-Net achieves promising accuracy in recognizing six distinct diver activities, outperforming state-of-the-art models. While the dataset provides a critical baseline, the work serves as a pioneering step toward smarter, collaborative underwater robotic systems.

*Auto-collected on 2026-06-12*