Event
QUAN: Quaternion Approximation Networks for Rotation-Equivariant Vision
2 reports · 2 sources
Bryce Grant and Peng Wang propose Quaternion Approximation Networks (QUAN), which use Hamilton product decomposition to approximate quaternion convolution as real-valued operations, paired with custom CUDA kernels, for rotation-equivariant image classification and object detection. The authors claim this approach preserves geometric properties while being easy to implement efficiently. To stabilize training, the authors also propose Independent Quaternion Batch Normalization (IQBN) and extend quaternion operations to spatial attention mechanisms. The paper evaluates QUAN on CIFAR-10/100 and ImageNet classification, COCO and DOTA detection, and robotic perception tasks. These results are as stated by the paper's authors and await independent verification.
Report timeline
- QUAN: Quaternion Approximate Networks Improve Image Classification and Oriented Object Detection
The paper proposes Quaternion Approximate Networks (QUAN), which use Hamilton product decomposition to approximate quaternion convolution with real-valued operations, paired with custom CUDA kernels to achieve rotation-equivariant image classification and object detection. The authors also propose Independent Quaternion Batch Normalization (IQBN) to stabilize training and extend quaternion operations to spatial attention mechanisms.
arXiv cs.RO ·
- QUAN: Quaternion Approximate Networks Improve Image Classification and Oriented Object Detection
The paper proposes Quaternion Approximate Networks (QUAN), which approximate quaternion convolution through Hamilton product decomposition using real-valued operations, paired with custom CUDA kernels to achieve rotation-equivariant image classification and object detection. The authors also propose Independent Quaternion Batch Normalization (IQBN) to stabilize training and extend quaternion operations to spatial attention mechanisms.
HuggingFace Daily Papers ·