Fully convolutional networks have been proved to be a powerful classifier being able to make inference on every single pixel of the input image. One of the key elements of such networks is the up-sampling procedure in the decoder of the network where the coarse output is connected with the dense pixels by using interpolation kernels in several layers of the decoder. Depending on the selected kernel, the final segmentation accuracy is affected even for fixed weight kernels. In this paper, various interpolation kernels are evaluated against network accuracy and time performance in several robot vision tasks. The performed quantitative assessments provide several insights for the best trade-offs between competitive performance and inference time.


A. Amanatiadis, V. Kaburlasos, E. Kosmatopoulos, “Interpolation kernels in fully convolutional networks and their effect in robot vision tasks”, Proceedings of the 2018 IEEE International Workshop on Imaging Systems & Techniques (IST 2018), Kraków, Poland, 16-18 October 2018, pp. 232-236.

View on ResearchGate