Skip to main content

Joel Douglas

Systems & Technology Research

Thursday, May 9, 2019
2PM – 3PM – HEC 101


State-of-the-art performance for object detection, object classification and activity recognition requires curation of large datasets which are necessary to make deep networks with tens of millions of parameters able to generalize. However, data of this scale is not available for many applications, especially in the infrared (IR) spectrum. For example, thermal IR is particularly difficult because the visible-spectrum convolutional networks are ill adapted to low-level spectral differences and many transfer learning algorithms will not improve accuracy. In this talk, we present Coupled-Instances Domain Adaptation (CIDA), which performs domain adaptation of a pre-trained convolutional network from a large source domain to a limited data target domain, by utilizing a small number of instances captured simultaneously in both domains. We collected a dataset of coupled instances of visible and thermal infrared of 98 subjects under various poses, designed for the training requirements of CIDA. We applied CIDA to the problem of cross-spectrum facial identification, where we achieved state-of-the-art performance in matching thermal infrared facial images to a visible mugshot gallery.