Self-Supervised Multi-Frame Monocular Scene Flow

CVPR 2021  ·  Junhwa Hur, Stefan Roth ·

Estimating 3D scene flow from a sequence of monocular images has been gaining increased attention due to the simple, economical capture setup. Owing to the severe ill-posedness of the problem, the accuracy of current methods has been limited, especially that of efficient, real-time approaches. In this paper, we introduce a multi-frame monocular scene flow network based on self-supervised learning, improving the accuracy over previous networks while retaining real-time efficiency. Based on an advanced two-frame baseline with a split-decoder design, we propose (i) a multi-frame model using a triple frame input and convolutional LSTM connections, (ii) an occlusion-aware census loss for better accuracy, and (iii) a gradient detaching strategy to improve training stability. On the KITTI dataset, we observe state-of-the-art accuracy among monocular scene flow methods based on self-supervised learning.

PDF Abstract CVPR 2021 PDF CVPR 2021 Abstract

Datasets


Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Scene Flow Estimation KITTI 2015 Scene Flow Test Multi-Mono-SF D1-all 30.78 # 2
D2-all 34.41 # 2
Fl-all 19.54 # 2
SF-all 44.04 # 3
Runtime (s) 0.063 # 1
Scene Flow Estimation KITTI 2015 Scene Flow Training Multi-Mono-SF D1-all 27.33 # 3
D2-all 30.44 # 1
Fl-all 18.92 # 1
SF-all 39.82 # 1
Runtime (s) 0.063 # 3

Methods