Adaptive deformable convolutional network

摘要

Deformable Convolutional Networks (DCNs) are proposed to solve the inherent limited geometric transformation in CNNs, showing outstanding performance on sophisticated computer vision tasks. Though they can rule out irrelevant image content and focus on region of interest to some degree, the adaptive learning of the deformation is still limited. In this paper, we delve it from the aspects of deformable modules and deformable organizations to extend the scope of deformation ability. Concretely, on the one hand, we reformulate the deformable convolution and RoIpooling by reconsidering spatial-wise attention, channel-wise attention and spatial-channel interdependency, to improve the single convolution's ability to focus on pertinent image contents. On the other hand, an empirical study is conducted on various and general arrangements of deformable convolutions (e.g., connection type) in DCNs. Especially on semantic segmentation, the study yields significant findings for a proper combination of deformable convolutions. To verify the effectiveness and superiority of our proposed deformable modules, we also provide extensive ablation study for them and compare them with other previous versions. With the proposed contribution, our refined Deformable ConvNets achieve state-of-the-art performance on two semantic segmentation benchmarks (PASCAL VOC 2012 and Cityscapes) and an object detection benchmark (MS COCO). (c) 2020 Elsevier B.V. All rights reserved.