Lightweight Monocular Depth with a Novel Neural Architecture Search Method

Abstract

This paper presents a novel neural architecture search method, called LiDNAS, for generating lightweight monocular depth estimation models. Unlike previous neural architecture search (NAS) approaches, where finding optimized networks is computationally highly demanding, the introduced novel Assisted Tabu Search leads to efficient architecture exploration. Moreover, we construct the search space on a pre-defined backbone network to balance layer diversity and search space size. The LiDNAS method outperforms the state-of-the-art NAS approach, proposed for disparity and depth estimation, in terms of search efficiency and output model performance. The LiDNAS optimized models achieve result superior to compact depth estimation state-of-the-art on NYU-Depth-v2, KITTI, and Scan-Net, while being 7%-500% more compact in size, i.e the number of model parameters.

Fast Neural Architecture Search for Lightweight Dense Prediction Networks

“…• We propose a multi-objective exploration framework, LDP, searching for accurate and lightweight dense prediction architectures. It extends our previous work [51] on monocular depth estimation to multiple prediction problems.…”

“…Moreover, the proposed method improves REL and RMSE by 22.3% and 18.7% while using only 3% of the model parameters comparing to the state-of-the-art NAS-based disparity and depth estimation approaches [86]. In addition, our method requires 90% less search time than [86] and outperforms [51] in almost all metrics.…”

“…Tables 10 and 11 show that our models produce more competitive results than several state-of-theart image super-resolution methods while being relatively compact. Our models perform on par with FPNet while be- Architecture CPU(ms) Ef+FBNet [94,97] 852 FSRCNN [24] 789 FastDepth [96] 458 VDSR [56] 425 PyD-Net [80] 226 Lite-HRNet [104] 217 LiDNAS-K [51] 205 LDP-Seg-Ci ing at least 32% smaller in terms of network size. Figure 10 provides visual comparisons on BSD100 and Urban100 benchmarks.…”

“…Third, most NAS studies tied architecture search to a single task, either semantic segmentation [16] or depth estimation [86,51] or image super-resolution [30]. By utilizing ATS together with a generic pre-defined backbone, LDP generalizes architecture search for various dense prediction problems.…”

“…Comparison on the NYU test set. (a) input image, (b) ground truth, (c) LDP-Depth-N, (d) LiDNAS-N[51], (e) Ef+FBNet[94,97], and (f) FastDepth[96].tern analysis and machine intelligence, 39(12):2481-2495, 2017. 3 [7] Bowen Baker, Otkrist Gupta, Nikhil Naik, and Ramesh Raskar.…”