This is a PyTorch implementation of horizontal circular convolution adaptable to panoramic images.
The figures below visualize gradients in 16x16 image space after proceeding multiple convolutions. The input image and convolution filters are initialized with a constant value of 1.
python main.py
Normal convolution with zero padding | Horizontal circular convolution |
---|---|
![]() |
![]() |
In our paper [1], we adapted this horizontal circular convolution on pseudo depth/reflectance images generated from 3D LiDAR dataset [2].
-
Learning Geometric and Photometric Features from Panoramic LiDAR Scans for Outdoor Place Categorization
K. Nakashima, H. Jung, Y. Oto, Y. Iwashita, R. Kurazume, O. M. Mozos
Advanced Robotics, Vol.32, No.14, 2018 -
Multi-modal Panoramic 3D Outdoor Datasets for Place Categorization
H. Jung, Y. Oto, O. Mozos, Y. Iwashita, R. Kurazume
In International Conference on Intelligent Robots and Systems (IROS), 2016