Balancing Calibration and Performance: Stochastic Depth in Segmentation BNNs


Linghong Yao (InstaDeep), Denis Hadjivelichkov (University College London), Andromachi Maria Delfaki (University College London), Yuanchang Liu (), Brooks Paige (University College London), Dimitrios Kanoulas (University College London)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

In many safety-critical applications, it is critical for computer vision models to provide reliable uncertainty estimates. However, traditional Bayesian approaches often compromise between efficiency and safety. In this work, we introduce a novel implementation of stochastic depth within segmentation Bayesian Neural Networks (BNNs) that preserves performance while significantly improving uncertainty calibration. We experimentally validate our approach using an encoder-decoder model specifically tailored for real-time robotic vision tasks which demand fast and reliable decision-making under inherently uncertain conditions. Our method facilitates both safer and more effective deployment without compromises, increasing uncertainty calibration error whilst maintaining high performance.

Citation

@inproceedings{Yao_2024_BMVC,
author    = {Linghong Yao and Denis Hadjivelichkov and Andromachi Maria Delfaki and Yuanchang Liu and Brooks Paige and Dimitrios Kanoulas},
title     = {Balancing Calibration and Performance: Stochastic Depth in Segmentation BNNs},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0546.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection