Instance Segmentation Based on Improved YOLACT
Lingyu Li, Ming Fang, Feiran Fu · 2020
YOLACT is the first real-time algorithm in instance segmentation. Aiming at the limitations in the algorithm such as low accuracy and poor stability for overlapping objects, an improved instance segmentation approach based on YOLACT was proposed. Specifically, the receptive field block is employed in the backbone network to extract multi-scale local region features, so that the network can learn more discriminative and robust features. To evaluate the validity of algorithm, experiments are performed on Pascal 2012 SBD. The results indicate the improved network can produce finer mask predictions in complex environment while keeping the real-time speed.