Conditional Convolutions for Instance Segmentation

  • 2020-03-12 08:42:36
  • Zhi Tian, Chunhua Shen, Hao Chen
  • 20

Abstract

We propose a simple yet effective instance segmentation framework, termedCondInst (conditional convolutions for instance segmentation). Top-performinginstance segmentation methods such as Mask R-CNN rely on ROI operations(typically ROIPool or ROIAlign) to obtain the final instance masks. Incontrast, we propose to solve instance segmentation from a new perspective.Instead of using instance-wise ROIs as inputs to a network of fixed weights, weemploy dynamic instance-aware networks, conditioned on instances. CondInstenjoys two advantages: 1) Instance segmentation is solved by a fullyconvolutional network, eliminating the need for ROI cropping and featurealignment. 2) Due to the much improved capacity of dynamically-generatedconditional convolutions, the mask head can be very compact (e.g., 3 conv.layers, each having only 8 channels), leading to significantly fasterinference. We demonstrate a simpler instance segmentation method that canachieve improved performance in both accuracy and inference speed. On the COCOdataset, we outperform a few recent methods including well-tuned Mask RCNNbaselines, without longer training schedules needed. Code is available: https://github.com/aim-uofa/adet

 

Quick Read (beta)

loading the full paper ...