Our paper, “SuperMETER: An Efficient Superpixel-Based Architecture for Monocular Depth Estimation on Resource-Constrained IoT Platforms,” has been accepted for publication in the IEEE Internet of Things Journal.
Abstract: This paper proposes SuperMETER, an efficient monocular depth estimation framework for resource-constrained IoT platforms. SuperMETER reformulates conventional pixel-wise dense prediction as superpixel-wise region-level sparse inference, reducing redundant computation, feature storage, and memory access while preserving structural information. It integrates convolution–superpixel fusion, superpixel-centered sparse regional decoding, structured skip fusion, and Gaussian kernel-based sparse-to-dense recovery, together with hardware–algorithm co-design for efficient fully on-chip FPGA deployment. On the NYU Depth v2 dataset, SuperMETER-XXS uses only 0.7M parameters and achieves an RMSE of 0.603 and a δ1 accuracy of 75.3%. Its implementation on the Xilinx ZCU102 achieves 179.1 FPS with fully on-chip execution and approximately 7× acceleration over existing methods.
Figure. Comparison of efficient monocular depth estimation paradigms. (a) Conventional pixel-wise dense prediction. (b) Proposed superpixel-wise sparse region inference for IoT edge platforms.