Matej Grcić presents Detecting anomalous pixels with or without negative data

Standard deep models cannot accommodate inputs that do not belong to the
training distribution. Hence, they often give rise to confident incorrect
predictions which may lead to devastating consequences. This problem is
especially demanding in the context of dense prediction since input images may
be only partially anomalous. We aim to mitigate this behaviour by complementing
semantic segmentation with anomaly detection to produce outlier-aware
segmentation outputs. We offer two design choices for dense anomaly detectors:
NFlowJS and DenseHybrid. NFlowJS is a divergence-based anomaly detector trained
on synthetic negative data produced by a normalizing flow. DenseHybrid fuses
data likelihood and dataset posterior into a hybrid anomaly detector. Both
approaches achieve strong performance on the standard benchmarks for anomaly
detection in road driving scenes. We also propose a novel performance metric
for
evaluating semantic segmentation in the open world. Our metric reveals a large
performance gap with respect to the closed-world setup, and thus contributes a
useful proxy towards advanced open-set recognition and safe autonomous driving.


Brief bio:
Matej Grcić received his MS degree from the University of Zagreb Faculty of
Electrical Engineering and Computing (UniZg-FER). He received a dean's award
for
outstanding individual research in 2020. He is currently pursuing his PhD
degree
at Uni-ZG FER. His research interests include generative modelling and robust
open-world dense recognition. His research has been accepted for presentation
at
VISAPP 2021, NeurIPS 2021 and ECCV 2022. He won the ACDC challenge for
semantic
segmentation in adverse driving conditions organized as part of the CVPR'22
workshop Vision for all seasons.


Matej is visiting VRG and will stay for the whole week (Mon-Fri). He will be
seated at G10.