Visual Semantic Context Encoding for Domain Prediction of Aircrafts

EasyChair Preprint 5096

Abstract

In existing CV works visual semantic context is often learned implicitly - this work uses an explicit representation instead and makes two distinct contributions: Firstly, it is shown that during data aggregation context can be used to remove irrelevant images. Secondly, extending the idea of context across multiple images, objects can be observed in characteristic domains. An original baseline, supervised CNNs and unsupervised mixture models are used to predict domains of airplanes. A CNN achieves the best classification performance with accuracies from 69% to 85% depending on the dataset variation. The entire framework can be applied to predict arbitrary domains of objects and provide a higher-level sense of scene understanding.

Keyphrases: Aerial data, Domain Prediction, computer vision, visual semantic context

@booklet{EasyChair:5096,
  author    = {Andreas Kriegler and Daniel Steininger and Wilfried Wöber},
  title     = {Visual Semantic Context Encoding for Domain Prediction of Aircrafts},
  howpublished = {EasyChair Preprint 5096},
  year      = {EasyChair, 2021}}