R(Det)^2: Randomized Decision Routing for Object Detection

TL;DR

This paper introduces R(Det)^2, a novel decision head for object detection that combines decision trees with neural networks, improving detection accuracy by 1.4-3.6% AP on MS-COCO.

Contribution

It proposes a new randomized decision routing method integrating decision trees into neural networks for enhanced object detection performance.

Findings

Achieves 1.4-3.6% AP improvement on MS-COCO

Effectively boosts feature learning and decision-making

Demonstrates the effectiveness of the proposed approach

Abstract

In the paradigm of object detection, the decision head is an important part, which affects detection performance significantly. Yet how to design a high-performance decision head remains to be an open issue. In this paper, we propose a novel approach to combine decision trees and deep neural networks in an end-to-end learning manner for object detection. First, we disentangle the decision choices and prediction values by plugging soft decision trees into neural networks. To facilitate effective learning, we propose randomized decision routing with node selective and associative losses, which can boost the feature representative learning and network decision simultaneously. Second, we develop the decision head for object detection with narrow branches to generate the routing probabilities and masks, for the purpose of obtaining divergent decisions from different nodes. We name this approach as the randomized decision routing for object detection, abbreviated as R(Det)$^2$. Experiments on MS-COCO dataset demonstrate that R(Det)$^2$ is effective to improve the detection performance. Equipped with existing detectors, it achieves $1.4\sim 3.6$\% AP improvement.