Cooperative Training of Triplet Networks for Cross-Domain Matching

  • Giovanni De Giacomo FURG
  • Matheus dos Santos FURG
  • Paulo Drews-Jr FURG
  • Silvia Botelho FURG

Resumo


Recently, Deep Convolutional Neural Networks have been applied to various computer vision problems and achieved state-of-the-art results. Among these, Siamese and Triplet networks have obtained great traction in intra-domain matching. However, is impossible to directly use these networks in cross-domain problems. Thus, this paper proposes a data-driven approach for cross-domain matching of complex data that do not share similar features. A pair of triplet networks are trained with a new cooperative approach to perform Deep Metric Learning. In order to validate our proposed method, we apply it to a cross-domain image matching problem that aims to assist with underwater robot localization. We train a pair of networks using our methodology on a dataset composed of acoustic and segmented aerial images and evaluate it on a dataset acquired in another location. Our results show that our method is able to achieve up to 83% accuracy in matching acoustic and segmented aerial images.
Palavras-chave: Training, Image segmentation, Sonar, Neural networks, Acoustics, Convolution, Pipelines
Publicado
09/11/2020
DE GIACOMO, Giovanni; DOS SANTOS, Matheus; DREWS-JR, Paulo; BOTELHO, Silvia. Cooperative Training of Triplet Networks for Cross-Domain Matching. In: SIMPÓSIO BRASILEIRO DE ROBÓTICA E SIMPÓSIO LATINO AMERICANO DE ROBÓTICA (SBR/LARS), 17. , 2020, Natal. Anais [...]. Porto Alegre: Sociedade Brasileira de Computação, 2020 . p. 192-197.