Learnable Polyphase Sampling for Shift Invariant and Equivariant Convolutional Networks
Learnable Polyphase Sampling for Shift Invariant and Equivariant Convolutional Networks
复制标题
DOI：
10.48550/arxiv.2210.08001
复制
发表时间：
2022-10
期刊：
影响因子：
--
通讯作者：
Renan A. Rojas-Gomez;Teck-Yian Lim;A. Schwing;M. Do;Raymond A. Yeh
中科院分区：
文献类型：
--
作者：
Renan A. Rojas-Gomez;Teck-Yian Lim;A. Schwing;M. Do;Raymond A. Yeh
我们提出了可学习的多相采样（LPS），一对可学习的下/上采样层，可实现真正的平移不变和等变卷积网络。 LPS 可以根据数据进行端到端训练，并概括现有的手工下采样层。它具有广泛的适用性，因为它可以通过替换下/上采样层集成到任何卷积网络中。我们在图像分类和语义分割方面评估 LPS。实验表明，LPS 在性能和班次一致性方面与现有方法相当或优于现有方法。我们首次在语义分割 (PASCAL VOC) 上实现了真正的平移等变，即 100% 平移一致性，绝对优于基线 3.3%。
We propose learnable polyphase sampling (LPS), a pair of learnable down/upsampling layers that enable truly shift-invariant and equivariant convolutional networks. LPS can be trained end-to-end from data and generalizes existing handcrafted downsampling layers. It is widely applicable as it can be integrated into any convolutional network by replacing down/upsampling layers. We evaluate LPS on image classification and semantic segmentation. Experiments show that LPS is on-par with or outperforms existing methods in both performance and shift consistency. For the first time, we achieve true shift-equivariance on semantic segmentation (PASCAL VOC), i.e., 100% shift consistency, outperforming baselines by an absolute 3.3%.