Research map / Architectures and efficiency / CNN design
Learning Spherical Convolution for Fast Features from 360° Imagery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt convolution kernels to equirectangular distortion and tie weights only within each image row | VGG feature extraction on spherical imagery · Pano2Vid test | conv5_3 output error | Helped |
| adapt convolution kernels to equirectangular distortion and tie weights only within each image row | VGG feature extraction on spherical imagery · PASCAL VOC 2007 test, transformed to 360° | multi-class classification accuracy | Helped |
| replace dense exact reprojection with SphConv | VGG feature extraction on spherical imagery · Pano2Vid test | MAC computation cost: about 400 times slower | Helped |
| replace Interp with SphConv | VGG feature extraction on spherical imagery · Pano2Vid test | MAC computation cost: 34% faster | Helped |
| replace Interp with SphConv | VGG feature extraction on spherical imagery · Pano2Vid test | feature output accuracy | Helped |
| use perspective cube-map projections instead of SphConv | VGG feature extraction on spherical imagery · Pano2Vid test | feature output accuracy | Hurt |
| remove joint fine-tuning after kernel-wise pre-training (ablation) | SphConv · Pano2Vid test | conv5_3 output error | Hurt |
| remove joint fine-tuning after kernel-wise pre-training (ablation) | SphConv · PASCAL VOC 2007 test, transformed to 360° | multi-class classification accuracy | Mixed |
| use exact preceding-layer outputs to avoid accumulated errors (ablation) | SphConv · PASCAL VOC 2007 test, transformed to 360° | multi-class classification accuracy | Helped |
| jointly fine-tune across layers and kernels (ablation) | SphConv initialized with kernel-wise pre-training · Pano2Vid test | conv5_3 output error | Helped |
| jointly fine-tune across layers and kernels (ablation) | SphConv initialized with kernel-wise pre-training · PASCAL VOC 2007 test, transformed to 360° | multi-class classification accuracy | Mixed |
| apply Direct convolutions to the equirectangular image rather than distortion-adapted kernels | VGG feature extraction on spherical imagery · Pano2Vid test | conv5_3 output error | Hurt |
Closest papers by meaning
- Unified Spherical Frontend: Learning Rotation-Equivariant Representations of Spherical Images from Any Camera (CVPR 2026 · similarity 0.90)
- Kernel Transformer Networks for Compact Spherical Convolution (CVPR 2019 · similarity 0.89)
- Tangent Images for Mitigating Spherical Distortion (CVPR 2020 · similarity 0.88)
- SpherePHD: Applying CNNs on a Spherical PolyHeDron Representation of 360deg Images (CVPR 2019 · similarity 0.88)
- DeepSphere: a graph-based spherical CNN (ICLR 2020 · similarity 0.87)
- SphereUFormer: A U-Shaped Transformer for Spherical 360 Perception (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.