Toward Efficient Simultaneous Detection and Segmentation

Chong Zhang, Zongxian Li, Qiong Liu, Yonghong Tian, Wei Zeng, Yaowei Wang, Wenbai Chen

Abstract: To solve the low-speed problem of two-stage based framework for object detection and instance segmentation, we creatively introduce the large separated convolution to the typical two-stage method. In our method, the two-branches separated large kernel convolution operation is applied before the ROI pooling layer, which is able to reduce the complexity of the follow-up process to a great extent and make the ROI pooling much more efficient. Furthermore, the subnet of region-based convolution network is carefully simplified and designed for obtaining better performances. Extensive evaluation experiments on Microsoft COCO datasets show that our method provides ∽2x speedup compared with the original Mask R-CNN method and results in a comparable detection and segmentation performances.