HMS-Net: Hierarchical Multi-scale Sparsity-invariant Network for Sparse Depth Completion

  • 2020-02-20 18:19:25
  • Zixuan Huang, Junming Fan, Shenggan Cheng, Shuai Yi, Xiaogang Wang, Hongsheng Li
  • 0

Abstract

Dense depth cues are important and have wide applications in various computervision tasks. In autonomous driving, LIDAR sensors are adopted to acquire depthmeasurements around the vehicle to perceive the surrounding environments.However, depth maps obtained by LIDAR are generally sparse because of itshardware limitation. The task of depth completion attracts increasingattention, which aims at generating a dense depth map from an input sparsedepth map. To effectively utilize multi-scale features, we propose three novelsparsity-invariant operations, based on which, a sparsity-invariant multi-scaleencoder-decoder network (HMS-Net) for handling sparse inputs and sparse featuremaps is also proposed. Additional RGB features could be incorporated to furtherimprove the depth completion performance. Our extensive experiments andcomponent analysis on two public benchmarks, KITTI depth completion benchmarkand NYU-depth-v2 dataset, demonstrate the effectiveness of the proposedapproach. As of Aug. 12th, 2018, on KITTI depth completion leaderboard, ourproposed model without RGB guidance ranks first among all peer-reviewed methodswithout using RGB information, and our model with RGB guidance ranks secondamong all RGB-guided methods.

 

Quick Read (beta)

loading the full paper ...