[Paper Review] R(Det)^2: Randomized Decision Routing for Object Detection

Yali Li, Shengjin Wang|arXiv (Cornell University)|Apr 2, 2022

Advanced Neural Network Applications5 citations

TL;DR

This paper proposes R(Det)², a novel end-to-end object detection framework that integrates soft decision trees into the decision head to disentangle routing decisions from prediction values. By introducing randomized decision routing with node-selective and associative losses, it enables divergent, representative feature learning across multiple tree nodes, achieving 1.4–3.6% AP gains on MS-COCO with existing detectors.

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.

Motivation & Objective

• To address the limitation of single-node decision heads in deep object detectors, which may restrict feature exploration and performance.
• To design a learnable, tree-based decision mechanism that separates routing decisions from prediction values for improved generalization.
• To enable multi-node, divergent decision-making in object detection through end-to-end training with soft decision trees.
• To improve detection accuracy—especially for large objects—by enhancing feature representation and decision diversity.

Proposed method

• Integrates soft decision trees into the decision head of object detectors to decouple routing decisions from prediction values.
• Proposes randomized decision routing using a narrow branch to generate routing probabilities and a wide branch to produce routing masks for node-level decisions.
• Introduces a node-selective loss to encourage diverse feature exploration across tree nodes and an associative loss to align predictions with ground-truth targets.
• Employs end-to-end training to jointly optimize the deep neural network and the soft decision tree components.
• Uses a tree-structured architecture where the final prediction is a weighted sum of node outputs based on routing probabilities.
• Applies the method to both one-stage and two-stage detectors (e.g., Faster R-CNN, Cascade R-CNN) with minimal architectural changes.

Experimental results

Research questions

• RQ1 Can a tree-based decision head with multi-node routing improve object detection performance compared to standard single-node heads?
• RQ2 How can soft decision trees be effectively integrated into deep neural networks for end-to-end training in object detection?
• RQ3 Does randomized decision routing with selective and associative losses enhance feature representation and decision diversity?
• RQ4 To what extent does R(Det)² improve detection accuracy across different object sizes and backbone architectures?

Key findings

• R(Det)² achieves a 1.4–3.6% AP improvement on MS-COCO when integrated into existing detectors like Faster R-CNN and Cascade R-CNN.
• With 12 epochs of training, R(Det)² achieves 50.0% AP on COCO using ResNeXt-101-DCN, outperforming Faster R-CNN and Cascade R-CNN.
• On the same backbone, R(Det)² achieves 54.1% AP and 72.4% AP₅₀ with 24 epochs and test-time augmentation, surpassing DyHead by 0.3% in AP₅₀ and 1.0% in APₗ.
• Using Swin-L as backbone, R(Det)² reaches 57.4% AP with multi-scale testing and 12 epochs, demonstrating strong generalization across architectures.
• The method significantly improves detection accuracy for large objects, indicating enhanced feature utilization in the decision head.
• The ablation study confirms that randomized routing with node-selective and associative losses is critical for performance gains.