TakuNet

Abstract

TakuNet is a lightweight Convolutional Neural Network (CNN) specifically designed for real-time inference on embedded Unmanned Aerial Vehicle (UAV) systems, particularly in emergency response scenarios (Rossi et al., 2025). The architecture balances high accuracy with low computational complexity, enabling efficient deployment on edge devices with limited power and processing capabilities. Published at the 2025 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), TakuNet addresses the critical need for energy-efficient computer vision in autonomous search and rescue operations.

Methodology

The TakuNet architecture follows a modular design optimized for hardware acceleration. It is constructed using a 4-stage macro-architecture that progressively increases channel width while reducing spatial resolution.

Architecture Design

The network is composed of the following key components:

• Stem Block: The initial processing unit responsible for rapid spatial downsampling and preliminary feature extraction.
• TakuBlock: The core building block of the network. It utilizes depth-wise separable convolutions to drastically reduce the number of parameters and floating-point operations (FLOPs) compared to standard convolutions. Each block incorporates skip connections (residual learning) to facilitate gradient flow and prevent degradation during training.
• DownSampler: Specialized modules placed between stages to reduce feature map resolution efficiently.

Model Configuration

The specific configuration of TakuNet used in the publication is defined by the following hyperparameters:

• Depths: [5, 5, 5, 4] (Number of blocks in each of the 4 stages)
• Widths: [40, 80, 160, 240] (Number of channels in each stage)

This configuration was empirically selected to maximize the accuracy-latency trade-off on embedded accelerators.

Installation and Usage

The official implementation leverages Docker to ensure reproducibility across different hardware platforms.

Setup

# Clone the repository
git clone https://github.com/DanielRossi1/TakuNet.git

# Build the Docker container
cd TakuNet/docker
./build.sh

# Run the container
./run.sh

Training and Inference

The framework provides a unified interface for training and inference through the launch.sh script:

cd src
# Launch training with default configuration
./launch.sh TakuNet

For deployment on edge devices (e.g., NVIDIA Jetson Orin Nano), the model supports export to TensorRT engines for maximum performance.

Implementation Details

The project is implemented using PyTorch Lightning, ensuring code modularity, reproducibility, and scalable training.

• Training Strategy: The model is trained using the RMSprop optimizer coupled with a StepLR learning rate scheduler to ensure stable convergence.
• Deployment Pipeline: To achieve real-time performance on edge devices, the training pipeline includes export utilities for optimized inference engines:
  • NVIDIA Jetson Orin Nano: The model is converted to TensorRT engines, utilizing FP16 precision to leverage the Tensor Cores on the Orin architecture.
  • Raspberry Pi: For CPU-bound edge devices, the model is exported to ONNX format and executed via ONNX Runtime.

Experimental Results

TakuNet was evaluated on the AIDER (Aerial Image Dataset for Emergency Response) and AIDER-V2 datasets.

• Inference Speed: On the NVIDIA Jetson Orin Nano, TakuNet achieves an inference throughput of over 650 FPS (Frames Per Second) in FP16 mode.
• Efficiency: The model demonstrates a superior balance of accuracy and computational cost compared to general-purpose lightweight backbones like MobileNetV3 and EfficientNet-Lite when applied to the specific domain of aerial emergency imagery.

Resources