NeRD++: Improved 3D-mirror symmetry learning from a single image is a follow-up work over NeRD: Neural 3D Reflection Symmetry Detector, which aims to detect the dominant mirror symmetry plane from a single-view image. This work has been accepted at BMVC 2022.
Yancong Lin, and Silvia Laura Pintea, and Jan C. van Gemert
Vision Lab, Delft University of Technology, the Netherlands
Abstract: Many objects are naturally symmetric, and this symmetry can be exploited to infer unseen 3D properties from a single 2D image. Recently, NeRD is proposed for accurate 3D mirror plane estimation from a single image. Despite the unprecedented accuracy, it relies on large annotated datasets for training and suffers from slow inference. Here we aim to improve its data and compute efficiency. We do away with the computationally expensive 4D feature volumes and instead explicitly compute the feature correlation of the pixel correspondences across depth, thus creating a compact 3D volume. We also design multi-stage spherical convolutions to identify the optimal mirror plane on the hemisphere, whose inductive bias offers gains in data-efficiency. Experiments on both synthetic and real-world datasets show the benefit of our proposed changes for improved data efficiency and inference speed.
This repo is not yet complete, still in progress!
Please refer to the baseline work NeRD for details.
conda env create -f environment.yml
Train: python train.py -d 0 --identifier nerd++ config/config.yaml
Test: python eval.py -d 0 --output result.npz path/config.yaml path/checkpoint.pth.tar
To do
@inproceedings{lin2021symmetry,
author={Lin, Yancong and Pintea, Silvia L and van Gemert, Jan C},
title = {{NeRD++}: Improved 3D-mirror symmetry learning from a single image},
year = {2022},
booktitle = {BMVC},
}