Vote-Based-EVT

The official code for the repo: Open-set Recognition for the Detection of Unknown Classes Based on theCombination of Convolutional Neural Networks and Random Forests

Dependencies

• Python (>=3.6)
• scikit-learn (0.21.0)
• scipy (1.5.2)
• numpy (0.19.0)
• Matlab 2020a

Table of Contents

1. MNIST
2. CIFAR
3. Traffic Scenarios

General descriptiopn

For all the datasets, the classes are divided into known and unknown classes. The known classes are chosen randomly and the process is repeated 5 times. The Macro F-Score of the known classes and the unknown class is calculated for the 5 different known class sets. The CNN trained in a supervised fashion for feature extraction and is used for feature extraction. Followed by which the Voter-Based EVT Model is trained and unknown classes are detected

MNIST

Step 1: To train MNIST CNN and extract features run sh Scripts/MNIST_CNN_Extraction.sh and to train the vote based model run the script \Mnist\RF+EVT\Vote_Based_EVT.m is Matlab to train the Vote-Based EVT Model

Step 2: Evaulation can be done by using the script \Mnist\RF+EVT\VoteBasedEVTStat.py

CIFAR

Step 1: To train CIFAR CNN and extract features run sh Scripts\CIFAR_CNN_Extraction.sh and to train the vote based model run the script \Cifar\RF+EVT\Vote_Based_EVT.m is Matlab to train the Vote-Based EVT Model

Step 2: Evaulation can be done by using the script \Cifar\RF+EVT\VoteBasedEVTStat.py

Traffic Scenarios

The traffic scenarios are generated from the HighD Dataset [1]. Please fill in the forms to request access to the HighD Data from https://www.highd-dataset.com/.

Step 1: Generate scenario categories using the script \Traffic_Scenarios\highD_generate_scenarios.m, occupancy grids will be generated for the scenarios and saved for CNN+RF training

Step 2: Train the CNN and extract features for traffic scenarios using python \Traffic_Scenarios\RF+EVT\ScenarioBasic.py, followed by that to train the vote based model run the script \Traffic_Scenarios\RF+EVT\Vote_Based_EVT.m

Step 3: Evaulation can be done by using the script \Traffic_Scenarios\RF+EVT\VoteBasedEVTStat.py

Reference

[1] The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems, Krajewski et al., ITSC 2018