Unlocking Early-Exiting Semantic Segmentation with Branched Networks

Author(s)

Gilbert, Mateus S. ; Pacheco, Roberto G. ; Couto, Rodrigo S. ; de Campos, Marcello L. R. ; Campista, Miguel Elias M.

• Document type: Journal article
• Source: 2023 IEEE LATIN-AMERICAN CONFERENCE ON COMMUNICATIONS, LATINCOM; v. N/A, p. 6-pg., 2023-01-01.

Abstract

Early-exit Deep Neural Networks (DNNs) and DNNs partitioning are valuable options to ease DNN implementation for image classification in resource-constrained devices and latency-sensitive applications. By inserting side branches into the usual DNN architecture, the resulting model can perform early classification attempts, using features extracted in layers that come before the model's output. DNN partitioning spreads DNN layers subsets across devices, the edge, and the cloud, bringing branches closer to the device. Offloading to the cloud is limited to hard-to-classify data. Given the success in image classification, it is natural to try to replicate these results with semantic segmentation. In this direction, we propose Branched-DeepLabV3 (BranDeepLabV3) as an early attempt to breach the gap between early-exit DNNs and semantic segmentation. Our results show that, by adding side branches to a semantic segmentation DNN, we can deliver adequate outputs without traversing the whole network. Moreover, our qualitative results show the usefulness of early-exits in semantic segmentation, as even coarse segmentations can be good enough to supply time-sensitive applications. (AU)

• FAPESP's process: 15/24494-8 - Communications and processing of big data in cloud and fog computing
• Grantee: Nelson Luis Saldanha da Fonseca
• Support Opportunities: Research Projects - Thematic Grants