Enhancing geometric modeling in convolutional neural networks: limit deformable convolution

Abstract Convolutional neural networks (CNNs) are constrained in their capacity to model geometric transformations due to their fixed geometric structure. To overcome this problem, researchers introduce deformable convolution, which allows the convolution kernel to be deformable on the feature map. However, deformable convolution may introduce irrelevant contextual information during the learning process and thus affect the model performance. DCNv2 introduces a modulation mechanism to control the diffusion of the sampling points to control the degree of contribution of offsets through weights, but we find that such problems still exist in practical use. Therefore, we propose a new limit deformable convolution to address this problem, which enhances the model ability by adding adaptive limiting units to constrain the offsets and adjusts the weight constraints on the offsets to enhance the image-focusing ability. In the subsequent work, we perform lightweight work on the limit deformable convolution and design three kinds of LDBottleneck to adapt to different scenarios. The limit deformable network, equipped with the optimal LDBottleneck, demonstrated an improvement in mAP75 of 1.4% compared to DCNv1 and 1.1% compared to DCNv2 on the VOC2012+2007 dataset. Furthermore, on the CoCo2017 dataset, different backbones equipped with our limit deformable module achieved satisfactory results. The source code for this work is publicly available at https://github.com/1977245719/LDCN.