Source code for 2021 ICCV paper "In-the-Wild Single Camera 3D Reconstruction Through Moving Water Surfaces"

Last update: Dec 06, 2022

Overview

In-the-Wild Single Camera 3D Reconstruction
Through Moving Water Surfaces

This is the PyTorch implementation for 2021 ICCV paper "In-the-Wild Single Camera 3D Reconstruction Through Moving Water Surfaces"

Project Page | Paper | Supplemental Material

In-the-Wild Single Camera 3D Reconstruction Through Moving Water Surfaces
Jinhui Xiong, Wolfgang Heidrich
KAUST
ICCV 2021 (Oral)

We propose a differentiable framework to estimate underwater scene geometry along with the time-varying water surface. The inputs to our model are a video sequence captured by a fixed camera. Dense correspondence from each frame to a world reference frame (selected from the input sequences) is pre-computed, ensuring the reconstruction is performed in a unified coordinate system. We feed the flow fields, together with initialized water surfaces and scene geometry (all are initialized as planar surfaces), into the framework, which incorporates ray casting, Snell’s law and multi-view triangulation. The gradients of the specially designed losses with respect to water surfaces and scene geometry are back-propagated, and all parameters are simultaneously optimized. The final result is a quality reconstruction of the underwater scene, along with an estimate of the time-varying water-air interface. The data shown here was captured in a public fountain environment.

Prerequisite

The code was tested with python>=3.7 & PyTorch>=1.3 & cuda>=10.0 on Nvidia RTX 2080 Ti
Minor change on the code if there is compatibility issue. It needs around 10 GB GPU memory.

Setup

conda create -n moving_water python=3.7
conda activate moving_water

conda install pytorch torchvision -c pytorch
conda install -c conda-forge opencv scikit-image
conda install -c anaconda scipy

Run the code

Please go to example folder, download the cached coefficient matrices (there are three matrices for each example) and execute:

python3 run.py

Citation

@inproceedings{xiong2021inthewild,
  title={In-the-Wild Single Camera 3D Reconstruction Through Moving Water Surfaces},
  author={Jinhui Xiong and Wolfgang Heidrich},
  year={2021},
  booktitle={ICCV}
}

Contact

Please contact Jinhui Xiong [email protected] if you have any question or comment.

Source code for 2021 ICCV paper "In-the-Wild Single Camera 3D Reconstruction Through Moving Water Surfaces"

Related tags

Overview

In-the-Wild Single Camera 3D Reconstruction
Through Moving Water Surfaces

Project Page | Paper | Supplemental Material

Prerequisite

Setup

Run the code

Citation

Contact

Owner

Gradient representations in ReLU networks as similarity functions

Learning Domain Invariant Representations in Goal-conditioned Block MDPs

Tiny Object Detection in Aerial Images.

A generalist algorithm for cell and nucleus segmentation.

Self-supervised Augmentation Consistency for Adapting Semantic Segmentation (CVPR 2021)

TensorFlow implementation of Elastic Weight Consolidation

Download & Install mods for your favorit game with a few simple clicks

A blender add-on that automatically re-aligns wrong axis objects.

MlTr: Multi-label Classification with Transformer

SoK: Vehicle Orientation Representations for Deep Rotation Estimation

New approach to benchmark VQA models

UV matrix decompostion using movielens dataset

MEAL V2: Boosting Vanilla ResNet-50 to 80%+ Top-1 Accuracy on ImageNet without Tricks

C3D is a modified version of BVLC caffe to support 3D ConvNets.

Fast EMD for Python: a wrapper for Pele and Werman's C++ implementation of the Earth Mover's Distance metric

You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling

Indoor Panorama Planar 3D Reconstruction via Divide and Conquer

Deep learning for spiking neural networks

Powerful unsupervised domain adaptation method for dense retrieval.

Official implementation of "UCTransNet: Rethinking the Skip Connections in U-Net from a Channel-wise Perspective with Transformer"