Evaluating variational autoencoder methods for out-of-distribution detection in autonomous vehicles

In a safety-critical system like autonomous vehicles, it is essential to ensure that the observations shown are within the distribution of training data, otherwise they are called out-of-distribution (OOD). OOD detection is a fundamental problem that needs to be addressed to avoid errors in image re...

وصف كامل

محفوظ في:
التفاصيل البيبلوغرافية
المؤلف الرئيسي: Dinh, Phuc Hung
مؤلفون آخرون: Arvind Easwaran
التنسيق: Final Year Project
اللغة:English
منشور في: Nanyang Technological University 2023
الموضوعات:
الوصول للمادة أونلاين:https://hdl.handle.net/10356/166524
الوسوم: إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
المؤسسة: Nanyang Technological University
اللغة: English
الوصف
الملخص:In a safety-critical system like autonomous vehicles, it is essential to ensure that the observations shown are within the distribution of training data, otherwise they are called out-of-distribution (OOD). OOD detection is a fundamental problem that needs to be addressed to avoid errors in image recognition tasks, especially in real-time system. Variational autoencoder (VAE) has emerged as the most promising method to address this issue. Several modifications have been made to VAE to improve its performance, especially in terms of increasing disentanglement, yet no research has been done to evaluate its performance on OOD detection. In this research, four VAE variants were tested on a traffic dataset to see which one gives the best results. After which, the relationship between disentanglement and OOD detection is evaluated.