A re-implemented project which focus on dash-cam images according to the paper (Deep Outdoor Illumination Estimation [Hold-Geoffroy et al. CVPR 2017]) (https://arxiv.org/abs/1611.06403). This project is an end-to-end system that outputs corresponding sun position and physcial sky, camera parameters by inputing single dash-cam image.

You can download the weights from here and our dataset from here.

If you want to test your own image, run this command:

python inference.py --img_path <image-path> --pre-trained <weight-path>

You can generate the dataset and list by using generate_data.py and the data (360 panorama images seperated into test and train) which followed the format in GS_skymodel.csv.

After generating dataset, run command below for training:

python train.py

The trained weights will be stored as weights.pth .

Evaluate the trained model on the test dataset (data/test_list.csv), it will output the average error of each predictions.

python eval.py --pre-trained <weight-path>

• numpy
• skimage
• pytorch
• opencv-python
• progressbar