CROSR

PyTorch implementation for "Classification-Reconstruction Learning for Open-Set Recognition" CVPR 2019. It is important to note that this repository borrows heavily from the repo - https://github.com/abhijitbendale/OSDN as far as fitting weibull distribution and computing open max scores goes. It also borrows from the original repo for the paper - https://nae-lab.org/~rei/research/crosr (which is implemented in chainer). I think thsi repository is much cleaner and flexible to accomodate newer datasets and model architectures. Some part of code works on python 3 and 2 whereas the other portion only works on python 2.7. Please follow the following steps to successfully run the code.

Usage

1) Compiling LibMR

Compile LibMR and python interface to LibMR using following commands. For pythong interfaces to work, you would require Cython to be pre-installed on your machine

cd libMR/
chmod +x compile.sh
./compile.sh

2) Train the DHRNet

python train_net.py

3) Compute the activation vectors for images

python get_model_features.py

4) Compute the MAV (mean activation vector) for each class category

python MAV_Compute.py

5) Compute the distance scores for activation features of training set

python compute_distances.py

6) Fit Weibull distribution for each category and calculate openmax scores (Note that this code needs to be run in Python 2.7.)

python compute_openmax.py

Results

The AUROC score for CIFAR-10 with 6/4 split is 71.23.