NavOCR

A lightweight, navigation-oriented OCR framework.

It is designed for robotic navigation scenarios, where only navigation-relevant text should be detected, such as:

• Signboards
• Room numbers

while irrelevant text, such as advertisements or price tags, is ignored.

Key features

• Focuses on navigation-relevant text to reduce unnecessary information and improve OCR speed
• Supports both standalone and ROS 2 integration
• Optimized for CPU-first robotic platforms, achieving ~6 FPS on CPUs
• Supports PaddlePaddle for GPU environments

Overview

• navocr_standalone.py: Run detection + OCR on a single image or a directory
• navocr/ros_node.py: ROS 2 node entry point
• configs/navocr_onnx.params.yaml: ONNX detector + ONNX OCR config (default)
• configs/navocr_openvino.params.yaml: OpenVINO detector + OpenVINO OCR config
• configs/navocr_paddle.params.yaml: PaddleDetection detector + Paddle OCR config
• configs/navocr_pytorch.params.yaml: PyTorch detector + Paddle OCR config

Backend Composition

Model format	Runtime / engine	Hardware	Text detection	Text recognition	FPS
ONNX	ONNX Runtime	CPU / GPU	RT-DETRv4 (Fine-tuned)	PP-OCRv5	4.51
OpenVINO IR	OpenVINO Runtime	Intel CPU	RT-DETRv4 (Fine-tuned)	PP-OCRv5	6.07
Paddle model	Paddle Inference	CPU / GPU	PP-YOLOE (Fine-tuned)	PP-OCRv5	1.79
PyTorch	PyTorch	CPU / GPU	RT-DETRv4 (Fine-tuned)	PP-OCRv5	2.05

*All FPS was measured on 11th Gen Intel(R) Core(TM) i5-1135G7.

Installation

Download Model

ONNX, OpenVINO, PaddlePaddle models are included in this repository.

git clone git@github.com:kc-ml2/NavOCR.git

Python Environment Setup (recommended)

Using a venv keeps NavOCR's Python dependencies isolated from the system Python and avoids conflicts with colcon build.

python3 -m venv ~/.venvs/navocr
source ~/.venvs/navocr/bin/activate

pip install --upgrade pip
pip install colcon-common-extensions

For ONNX runtime (default)

pip install onnxruntime pyyaml opencv-python numpy

If you want to run ONNX Runtime on CUDA, install onnxruntime-gpu instead of onnxruntime.

For OpenVINO runtime (Optional)

pip install openvino pyyaml opencv-python numpy

For PyTorch runtime (Optional)

pip install pyyaml opencv-python numpy
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

# Download weight
pip install gdown==5.2.0
gdown https://drive.google.com/uc?id=1D0zRUmyPvxgXq2rytrzVqg3-5_-5gkDn
mv rtv4_hgnetv2_s.pth model/pytorch/rtv4_hgnetv2_s.pth

For Paddle runtime (Optional)

This is only required for paddlepaddle backend.

Tested with paddlepaddle==3.0.0 and paddleocr==3.4.0.

Install PaddlePaddle following the official installation guide for your OS / Python / CUDA version:

• https://www.paddlepaddle.org.cn/en/install/quick

Then install PaddleDetection and PaddleOCR:

pip install pyyaml opencv-python numpy

# PaddleDetection
git clone https://github.com/PaddlePaddle/PaddleDetection.git

cd PaddleDetection
pip install -r requirements.txt
python setup.py install

# PaddleOCR
pip install paddleocr

Standalone Inference

Download Testset

# Setup python env
pip install gdown==5.2.0

# Download sample testset
mkdir data && cd data
gdown https://drive.google.com/uc?id=1GcgddRm4GsjPKUOVdmWFzeF5gElCZfx2
unzip example_sequence.zip 
cd .. && mkdir results

Run with ONNX runtime (default)

python navocr_standalone.py \
  --params-file configs/navocr_onnx.params.yaml \
  --infer_dir data/example_sequence/images

Run with OpenVINO runtime

# If you encounter oneDNN compatibility issues on CPU, set these before running:
export FLAGS_enable_pir_api=0
export FLAGS_enable_pir_in_executor=0

python navocr_standalone.py \
  --params-file configs/navocr_openvino.params.yaml \
  --infer_dir data/example_sequence/images

Run with PyTorch runtime

python navocr_standalone.py \
  --params-file configs/navocr_pytorch.params.yaml \
  --infer_dir data/example_sequence/images

Run with Paddle runtime

python navocr_standalone.py \
  --params-file configs/navocr_paddle.params.yaml \
  --infer_dir data/example_sequence/images

Single image

python navocr_standalone.py \
  --params-file configs/navocr_onnx.params.yaml \
  --input data/example_sequence/images/000000.jpg

ROS 2 Node

Build ROS 2 package

ROS dependencies are declared in package.xml. Install them from the workspace root with:

cd ~/ros2_ws  # your ros2 workspace
rosdep install --from-paths src --ignore-src -r -y

colcon build --symlink-install --packages-select navocr
python -m colcon build --symlink-install --packages-select navocr  # if you're using venv

source install/setup.bash

Run ros2 node

ros2 run navocr navocr_with_ocr_node

The default ROS 2 params file is configs/navocr_onnx.params.yaml.

If you want to select a different params file at runtime:

ros2 run navocr navocr_with_ocr_node --ros-args \
  -p params_file:=/absolute/path/to/configs/navocr_pytorch.params.yaml

ros2 run navocr navocr_with_ocr_node --ros-args \
  -p params_file:=/absolute/path/to/configs/navocr_openvino.params.yaml

ros2 run navocr navocr_with_ocr_node --ros-args \
  -p params_file:=/absolute/path/to/configs/navocr_paddle.params.yaml

Published topics:

• detections_topic default: /navocr/detections
• annotated_image_topic default: /navocr/annotated_image

Acknowledgements

We gratefully acknowledge the open-source projects that made this work possible: RT-DETRv4, PaddleDetection, PaddleOCR / PP-OCRv5, OpenVINO, and ONNX.

🚧 Planned Updates

We're working on expanding support beyond store signboards detection model. Stay tuned for upcoming features for broader navigation use cases.

• Library migration due to a license issue (ultralytics -> PaddleDetection)
• Alternative inference for higher FPS on CPU (Add OpenVINO support)
• Integration with text recognition (PaddleOCR)
• Integration with SLAM packages via ROS (TextMap)
• Model training scripts (Dataset crawling, model fine-tuning, ...)
• Floor sign detection
• Directional guide text detection

License

This repository is licensed under the Apache License, Version 2.0.

This project includes code and configuration files derived from PaddleDetection (https://github.com/PaddlePaddle/PaddleDetection) and RT-DETRv4 (https://github.com/RT-DETRs/RT-DETRv4), which are also licensed under the Apache License, Version 2.0.