Multi-level OOD detection in image segmentation networks for safety in automotive applications

This research paper aims to study the efficacy of utilizing multi-layer outputs from semantic segmentation models for handling various types of out-of-distribution (OOD) samples. Specifically, our focus lies on the comparison between early and late layers to determine which layer proves more e...

Full description

Saved in:
Bibliographic Details
Main Author: Poh, Eugene Yang Quan
Other Authors: Arvind Easwaran
Format: Final Year Project
Language:English
Published: Nanyang Technological University 2024
Subjects:
Online Access:https://hdl.handle.net/10356/175033
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:This research paper aims to study the efficacy of utilizing multi-layer outputs from semantic segmentation models for handling various types of out-of-distribution (OOD) samples. Specifically, our focus lies on the comparison between early and late layers to determine which layer proves more effective in addressing the different kinds of OOD categories. In pursuit of this objective, we analyze the output generated by different layers of the semantic segmentation model and assess their influence on the OOD detection performance when confronted with diverse OOD scenarios. Our study encompasses an examination of two primary OOD types, namely samples with covariate shift (e.g., white noise) and semantically OOD samples. The outcomes of this research aim to contribute valuable insights into the optimal utilization of multi-layer information in semantic segmentation models for improved OOD handling. By discerning the strengths and weaknesses of early and late layers in the context of distinct OOD challenges, we seek to provide a nuanced understanding of the interplay between model architecture and OOD sample characteristics. Ultimately, we hope that our findings may pave the way for enhanced model robustness and generalization capabilities in real-world scenarios.