Kernel-aware dynamic convolution for dense prediction

Gaoge Han

doi:10.1016/j.patcog.2025.112131

Kernel-aware dynamic convolution for dense prediction

2025-07-11

0

PRE

AI

DOI：10.1016/j.patcog.2025.112131

Original

Original request for help

Save

Abstract

En 中文

Recent dynamic convolution methods present the dynamism over learnable weights, which have been proved to increase representation ability. However, their dynamic designs can only target a specific problem. In this paper, we propose a new convolution called Kernel-aware Dynamic Convolution (KDConv), which extends the dynamism over applied kernels to explicitly address both the common visual spatial geometric transformation and scale variation problems using an unified way. Specifically, KDConv adaptively determines the most suitable kernel from the filter pool in each spatial location through a kernel-aware module. In the filter pool, we set multiple shared filters with different size and offset parameters. The kernel-aware module builds the dependency between these kernels and generates the index map for selection. Our design of kernel-aware dynamic size and offset is specifically advantageous for dense prediction including object detection and instance segmentation. Extensive experiments show that KDConv achieves largely performance improvement over the state-of-the-art dynamic convolution methods.

Keywords:

dynamic convolution

kernel-aware module

filter pool

kernel selection

dense prediction

Kernel-aware dynamic convolution for dense prediction

Abstract

Journal

Organization