Crop & Match: RoI Cropping and Feature Matching for Segmentation of Small Objects
Goo-Young Moon, Jong‐Ok Kim · 2023
In semantic segmentation, predicting RoI (Region of Interest) classes or objects within small regions faces challenges. To address this issue, we propose Crop&Match, a method to combine RoI cropping and feature matching, to enhance model performance of small RoI classes. The cropping ensures the network to focus on small RoI regions by cropping image regions to contain RoI. And feature matching aligns input images with their RoI cropped versions at a representation level, improving predictive consistency across different perspectives on the same image, thereby boosting RoI class recognition.