New Transformer Model Achieves Efficient Edge Action Recognition

Researchers have developed CoDAT, a Collaborative Dual-Attention Transformer designed for efficient action recognition on edge devices. This model utilizes a lightweight dual-branch attention mechanism, combining Spatial Convolutional Attention (SCA) for local feature aggregation and Strided Single-Head Attention (SSHA) for global context, significantly reducing computational costs. CoDAT also incorporates a parameter-free TShift module for temporal modeling, enabling efficient communication across frames. Experiments show CoDAT achieves a superior energy-accuracy balance compared to existing methods on various benchmarks, offering faster throughput and fewer parameters for real-time applications in edge IoT systems. AI

IMPACT Enables more efficient and real-time AI-powered perception systems on resource-constrained edge devices.

RANK_REASON The item is a research paper detailing a new model architecture and its performance on benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Collaborative Dual-Attention Transformer
• EfficientViT384
• FastViT-S12
• ImageNet-1K
• Internet-of-Things
• Kinetics-400
• LAPS
• MA-52
• Jetson AGX Orin
• Raspberry Pi 5
• Spatial Convolutional Attention
• Strided Single-Head Attention
• TokShift
• TShift module
• UniFormer-B
• ViT-Temporal-Shift
• VSwin-T

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →