This repository contains code for estimating the direction of the sun from a single outdoor image.
In many areas of the world the sun shines a lot and humans have used it as a navigational aid for thousands of years. So maybe it would be fun to imbue our robotic creations with this ability too.
Applications:
- You can estimate the yaw angle between multiple cameras on the same rig even if they don't have overlap
- You can stabilise SLAM algorithms by adding an absolute yaw reference
- If you know the time, you know which way the sun is relative to you, and you have a reasonable estimate of where in the world you are then you can tell global heading
The algorithm doesn't work 🤷♂️
- pretrained_resnet.py: Defines a regression model based on a pre-trained ResNet model and includes functions for training and loading the model.
- 3_224_224_resnet.pth: A pre-trained ResNet model for estimating the direction of the sun from a single outdoor image.
- suncompass.py: Class for interacting with the pre-trained ResNet model.
- demo.py: Demonstrates how to use the pre-trained ResNet model to estimate the direction of the sun from a single outdoor image.
pip install suncompass-
Clone the repository:
git clone https://github.com/hugohadfield/suncompass.git cd suncompass -
Install the package:
pip install .
import numpy as np
import cv2
import matplotlib.pyplot as plt
from suncompass import SunCompass
image_path = "croatia.jpg"
suncompass = SunCompass()
suncompass.set_eval(dropout=False)
image = cv2.imread(image_path)
img_with_suncompass, theta_rad = suncompass.predict_and_draw(image)
plt.figure()
plt.imshow(cv2.cvtColor(img_with_suncompass, cv2.COLOR_BGR2RGB))
plt.title(f"Sun direction: {np.degrees(theta_rad):.1f} degrees")
plt.show()- Python 3.7+
- PyTorch
- torchvision
- pandas
- numpy
- matplotlib
- tqdm
- click
- PIL
- tensorboard
- suncalc
This script calculates the sun vector in the ENU frame and the relative frame based on heading.
Example usage:
python add_sun_vector.pyThis script extracts motion parameters such as speed, heading, and curvature from a given dataset.
Example usage:
python extract_motion.pyThis script defines and trains a regression model based on a pre-trained ResNet model.
Example usage:
python pretrained_resnet.pyThis script runs the trained ResNet model on a set of images and saves the results to a CSV file.
Example usage:
python run_model_on_images.py --model_name 3_224_224_resnet.pth --image_dir /path/to/images --gt_csv /path/to/ground_truth.csvcalculate_sun_vector_ENU(time: datetime, lat_degs: float, long_degs: float) -> np.ndarray: Calculates the sun vector in the ENU frame.calculate_sun_vector_relative(time_unixus: int, lat_degs: float, long_degs: float, heading_rads_from_east: Union[float, np.complex128]) -> np.ndarray: Calculates the sun vector in the relative frame.calculate_relative_azimuth_angle(time_unixus: int, lat_degs: float, long_degs: float, heading_rads_from_east: float) -> float: Calculates the azimuth angle of the sun vector in the relative frame.
haversine(lat1, lon1, lat2, lon2) -> float: Calculates the great circle distance between two points on the Earth.calculate_bearing_clockwise_from_north(lat1, lon1, lat2, lon2) -> float: Calculates the bearing between two points.calculate_heading_anticlockwise_from_east(lat1, lon1, lat2, lon2) -> float: Calculates the heading anti-clockwise from east.extract_motion(data: pd.DataFrame) -> pd.DataFrame: Extracts speed, heading, and curvature from the given data.
ResNetRegression: Defines a regression model based on a pre-trained ResNet model.train_network(device, n_epochs: int = 10, image_size: Tuple[int, int] = (128, 128)): Trains the network for a set number of epochs.
run_model_on_images(model_name: Path, image_dir: Path, image_size: Tuple[int, int, int] = (1, 100, 100)): Runs the model on the images and saves the results to a CSV file.plot_results_csv(gt_csv: Optional[str] = None): Plots the results from the CSV file.
This project is licensed under the MIT License - see the licence.txt file for details.