Decision-Making for Multi-View Single Object Detection with Graph Convolutional Networks

Aggregating predicted outputs from multiple views helps boost multi-view single object detection performance. Decision-making strategies are flexible to perform this result-level aggregation. However, the relationship among multiple views is not exploited in aggregation. This study proposes a novel decision-making model with graph convolutional networks (DM-GCN) to address this issue by establishing a relationship among predicted outputs with graph convolutional networks. Through training, the proposed DM-GCN learns to make a correct decision by enhancing the contributions from informative views. DM-GCN is light, fast, and can be applied to any object detector with a negligible computational cost. Moreover, a real captured dataset named Yogurt10 with a new metric is proposed to investigate the performance of DM-GCN in the multi-view single object detection task. Experimental results show that DM-GCN achieves superior performance compared to classical decision-making strategies. A visual explanation is also provided to interpret how DM-GCN makes a correct decision.