The scientific paper "Normalizing Flow-based Feature Synthesis for Outlier-Aware Object Detection" has been selected as one of the top 10% of papers (conference highlights) to be presented at the CVPR 2023 - Conference on Computer Vision and Pattern Recognition.
It is a collaboration between Professor Siniša Šegvić and researchers Nishant Kumar and Stefan Gumhold from the Technical University of Dresden and Abouzar Eslami from Carl Zeiss Meditec AG.
The article considers the rejection of detection responses in anomalous regions of test images. The presented approach is based on learning anomaly detectors on synthetic negative features. The presented experiments have shown that normalizing flows can generate much better synthetic negative features than classical generative models based on per-class Gaussian distributions. The full text of the article is available on Arxiv.
CVPR is the central annual conference in the field of computer vision and one of the most prestigious conferences in the field of technical sciences. This year the conference will be held from June 18 to 22, 2023, in Vancouver, Canada.
This year, CVPR accepted less than 26% of the received papers (a total of 9,155 papers were submitted), and only about 2.5% of the received papers entered the selection of the best articles at the conference.