👑 Input-Output-Conditioned Uncertainty Estimation (IO-CUE) 👑

Paper: https://arxiv.org/abs/2506.00918

A minimal implementation of the proposed IO-CUE framework, including all experiments shown in the submission.

Getting the data

This link provides the NYU Depth v2 and ApolloScape datasets as seen in Amini et al.: https://www.dropbox.com/s/qtab28cauzalqi7/depth_data.tar.gz?dl=1
Unpack the data to a convenient path and make sure to update paths in dataloaders/simple_depth.py accordingly

Train a base model using the train_base_model.py script.
Train a post-hoc learner using the train_post_hoc_model.py script and a config file. For example: python train_post_hoc_model.py -yc configs/yaml_configs/edgy_depth_gaussian_io_cue.yaml -d 0.

Name		Name	Last commit message	Last commit date
Latest commit History 77 Commits
configs		configs
dataloaders		dataloaders
evaluation		evaluation
intro_experiments		intro_experiments
models		models
networks		networks
predictors		predictors
probing_experiments		probing_experiments
.gitignore		.gitignore
LICENSE		LICENSE
MONOVIT_INTEGRATION.md		MONOVIT_INTEGRATION.md
README.md		README.md
create_metrics_table.py		create_metrics_table.py
requirements.txt		requirements.txt
run_uci_experiments.sh		run_uci_experiments.sh
setup_base_model.py		setup_base_model.py
test_monovit_integration.py		test_monovit_integration.py
test_uncertainty_augmentations.py		test_uncertainty_augmentations.py
train_base_model.py		train_base_model.py
train_post_hoc_model.py		train_post_hoc_model.py