BNN semantic segmentation based on work What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?
Pytorch implementation of "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?", NIPS 2017 This work use the repo as main skeleton but change the task from regression to semantic segmentation as the paper shows. You can learn more details by reading a nice blog.
Input
Results
(for ETH racing team AMZ, note that this is not public) Here we show the result from epoch 0 - 20 - 40 - 60 -80. You can see how the results change.
As you can see above the aleatoric is high around the cone boundary. This make sense because the boundary part is related to the measurement precision.
As for Epistemic uncertainty you can see it is larger where cone is far away. This indicates that we have more uncertainty in large range and need more data for that.
You can download the CamVid dataset by looking at this [repo]](https://github.com/alexgkendall/SegNet-Tutorial). You can also make your own dataset class easily follow the template.
# Project Tree
BNN_seg
├── BNN_seg_src/
│ ├── data/ *.py
│ ├── loss/ *.py
│ ├── model/ *.py
│ └── *.py
└── BNN_seg_exp/
├── log/
├── model/
└── save/
# Classification loss only
python main.py --uncertainty "normal"
# Epistemic / Aleatoric
python main.py --uncertainty ["epistemic", "aleatoric"]
# Epistemic + Aleatoric
python main.py --uncertainty "combined"
# Classification loss only
python main.py --is_train false --uncertainty "normal" --exp_load "the name of experiment siffix"
# Epistemic
python main.py --is_train false --uncertainty "epistemic" --n_samples 25 [or 5, 50] --exp_load "the name of experiment siffix"
# Aleatoric
python main.py --is_train false --uncertainty "aleatoric" --exp_load "the name of experiment siffix"
# Epistemic + Aleatoric
python main.py --is_train false --uncertainty "combined" --n_samples 25 [or 5, 50] --exp_load "the name of experiment siffix"
Note that if your model name is "combined_1027_1450" then for exp_load you should use 1027_1450
conda create --name BNN_Seg python=3.8
conda activate BNN_Seg
pip install -r requirement.txt
This is not official implementation.
-
Autoencoder based on Bayesian Segnet
- Network depth 3
- Drop_rate 0.5
Input
Results
Input
Results
Input
Results
Input
Results
Input
Results
Input
Results