/CSPN_monodepth

Unofficial Faster PyTorch implementation of Convolutional Spatial Propagation Network

Primary LanguagePython

CSPN implemented in Pytorch 0.4.1

Introduction

This is a PyTorch(0.4.1) implementation of Depth Estimation via Affinity Learned with Convolutional Spatial Propagation Network. At present, we can provide train script in NYU Depth V2 dataset for depth completion and monocular depth estimation. KITTI will be available soon!

Faster Implementation

We re-implement CSPN using Pixel-Adaptive Convolution.

Multi_GPU

The implementation of multi-gpus is based on inplace abn.

Results

Method Implementation details rml rmse log10 Delta1 Delta2 Delta3
Paper batch size=24 epoch=40 0.016 0.117 - 0.992 0.999 1.000
Our_impl batch size=8 iteration=100k 0.018 0.127 0.008 0.991 0.998 1.000
Our_CSPN batch size=8 iteration=100k 0.018 0.127 0.008 0.991 0.998 1.000

Image text

Third Libs

inplace abn

Pixel-Adaptive Convolution