Paper proposes Quaternion Approximate Networks for vision tasks
Original title: Quaternion Approximate Networks for Enhanced Image Classification and Oriented Object Detection
The authors report better accuracy with fewer parameters than prior convolution and quaternion models on standard vision benchmarks.
A paper introduces Quaternion Approximate Networks (QUAN), which approximate quaternion convolution through Hamilton product decomposition using real-valued operations and custom CUDA kernels. The design adds Independent Quaternion Batch Normalization and extends quaternion operations to spatial attention. Evaluations cover image classification on CIFAR-10/100 and ImageNet, object detection on COCO and DOTA, and robotic perception tasks. The authors report higher accuracy, fewer parameters, and faster convergence than existing convolution and quaternion models, and say QUAN sets the state of the art among quaternion CNNs for oriented detection.
- QUAN keeps geometric properties while using real-valued operations.
- Detection results emphasize parameter efficiency and rotation handling versus standard CNNs.
- The authors point to resource-constrained robotic perception as a target use.
- Code is linked from the paper page.
- AI models
- Quaternion Approximate Networks
From the original article
1 min read at huggingface.co
This paper introduces Quaternion Approximate Networks (QUAN), a novel deep learning framework that leverages quaternion algebra for rotation equivariant image classification and object detection. Unlike conventional quaternion neural networks attempting to operate entirely in the quaternion domain, QUAN approximates quaternion convolution through Hamilton product decomposition using real-valued operations.
By Bryce Grant, Peng Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.