dlwbm123/fvi_cv

Code for Scalable Uncertainty for Computer Vision with Functional Variational Inference @ CVPR 2020

Scalable Uncertainty for Computer Vision with Functional Variational Inference

Code for Scalable Uncertainty for Computer Visition with Functional Variational Inference, by Eduardo D C Carvalho, Ronald Clark, Andrea Nicastro and Paul H J Kelly. To appear at CVPR 2020.

Pre-requisites:

• Download folder /datasets, containing CamVid and pre-processed Make3D: https://drive.google.com/open?id=1HaEnG8smGgCzUpUUMnVxXInrBGh-rjyl
• Install cnn_gp package: https://github.com/cambridge-mlg/cnn-gp

Put /datasets folder at the same directory level as /FVI_CV.

Training all models:

Supply BASE_DIR as the path containing both /FVI_CV and /datasets folders.

If more convenient, set --n_epochs to a number smaller than 4000 and use --load in order to resume training.

Semantic segmentation on CamVid:

Ours-Boltzmann: python3 run_fvi_seg.py --base_dir=BASE_DIR --training_mode

Deterministic-Boltzmann/MCDropout-Boltzmann: python3 run_mcd_seg.py --base_dir=BASE_DIR --training_mode

Depth regression on Make3D:

Ours-Gaussian: python3 run_fvi_gaussian_depth.py --base_dir=BASE_DIR --training_mode

Ours-Laplace: python3 run_fvi_laplace_berhu_depth.py --base_dir=BASE_DIR --training_mode --likelihood="laplace"

Ours-berHu: python3 run_fvi_laplace_berhu_depth.py --base_dir=BASE_DIR --training_mode --likelihood="berhu"

Deterministic-Laplace: python3 run_deterministic_depth.py --base_dir=BASE_DIR --training_mode --loss="l1"

Deterministic-berHu: python3 run_deterministic_depth.py --base_dir=BASE_DIR --training_mode --loss="berhu"

MCDropout-Laplace: python3 run_mcd_laplace_depth.py --base_dir=BASE_DIR --training_mode

Computing test results:

Firstly, put all trained models (.bin files) and Ours-berHu likelihood threshold c_test.txt in folder /models_test. Change those filenames to the ones prescribed inside the if __name__ == '__main__': block in each run_*.py file.

Write same commands as in training, but replacing the --training_mode flag by --test_mode. For obtaining runtime comparison results, use --test_runtime_mode instead.

Finally, write python3 compare_calibration.py for computing the mean calibration scores.