Attention-based Dynamic Filters for Anchor-free Instance Segmentation

The convolution operation is the core of convolutional neural networks (CNNs). To make CNNs more efficient, existing works construct multi-scale representation by utilizing different filter sizes or expanding filter sizes with dilated convolutions. However, these filters have fixed parameters after training so that they are not adaptive to the input image during inference. To address this issue, we propose an attention-based dynamic filter, which is a novel design that adaptively generates filters based on image contents. We apply the proposed dynamic filter to the mask branch, named Attention-based Adaptive Con-guided mask (ACG-Mask) branch, which is added to anchor-free one-stage object detector (FCOS). Besides, we design a multi-scale head, which contains an improved Receptive Field Block (iRFB) to enhance the discriminability and robustness of the feature. We name our model as Attention-based Dynamic Filters for anchor-free Instance Segmentation (ADFInst). Extensive experiments on the fine-annotation Cityscapes and COCO datasets reveal the effectiveness of the proposed method. ADFInst achieves a new record 37.9% AP and 63.3% AP50 on the fine-annotation Cityscapes dataset and achieves 37.8% AP, 58.7% AP50, and 40.5% AP75 on COCO dataset.