MobileStereoNet: Towards Lightweight Deep Networks for Stereo Matching

22 Aug 2021  ·  Faranak Shamsafar, Samuel Woerz, Rafia Rahim, Andreas Zell ·

Recent methods in stereo matching have continuously improved the accuracy using deep models. This gain, however, is attained with a high increase in computation cost, such that the network may not fit even on a moderate GPU. This issue raises problems when the model needs to be deployed on resource-limited devices. For this, we propose two light models for stereo vision with reduced complexity and without sacrificing accuracy. Depending on the dimension of cost volume, we design a 2D and a 3D model with encoder-decoders built from 2D and 3D convolutions, respectively. To this end, we leverage 2D MobileNet blocks and extend them to 3D for stereo vision application. Besides, a new cost volume is proposed to boost the accuracy of the 2D model, making it performing close to 3D networks. Experiments show that the proposed 2D/3D networks effectively reduce the computational expense (27%/95% and 72%/38% fewer parameters/operations in 2D and 3D models, respectively) while upholding the accuracy. Our code is available at https://github.com/cogsys-tuebingen/mobilestereonet.

PDF Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Depth Estimation 2D-MobileStereoNet MACs 68.86 # 1
Stereo Depth Estimation KITTI2015 2D-MobileStereoNet three pixel error 2.67 # 3
Stereo Depth Estimation KITTI2015 3D-MobileStereoNet three pixel error 1.69 # 1
Stereo Depth Estimation sceneflow 3D-MobileStereoNet Average End-Point Error 0.80 # 1
EPE 0.80 # 1
Stereo Depth Estimation sceneflow 2D-MobileStereoNet Average End-Point Error 1.14 # 3
EPE 1.14 # 2

Methods