Abstract
Modern object detection frameworks have traditionally focused on backbone architectures for feature extraction, often utilizing simplistic, single-node linear layers for the final decision head. This paper introduces R(Det)², a novel decision head that replaces standard linear layers with a Soft Decision Tree structure. By employing randomized routing and specialized loss functions—Selective and Associative losses—the proposed method forces different nodes within the tree to learn diverse visual cues. Evaluated on the MS-COCO 2017 dataset, R(Det)² demonstrates significant performance gains across various architectures, including a 3.6% AP increase for Faster R-CNN and a state-of-the-art 57.4% AP when combined with a Swin-Large backbone.
Introduction
In the evolution of object detection, the "backbone" (responsible for feature extraction) and the "neck" (responsible for feature fusion) have seen immense innovation. However, the "head"—the component that translates these features into class labels and bounding box coordinates—has remained relatively stagnant. Most state-of-the-art detectors rely on a "single-node" approach, consisting of a few stacked fully connected or convolutional layers.
This simple design creates a bottleneck, limiting the network's ability to explore complex and diverse visual evidence. While previous efforts have attempted to disentangle classification and localization tasks, the underlying decision mechanism has remained a "black box" of linear layers. R(Det)² addresses this by integrating the hierarchical logic of decision trees directly into the end-to-end learning process of deep neural networks, allowing for more nuanced feature interpretation.
Methodology
The core innovation of R(Det)² is the replacement of standard prediction layers with a Soft Decision Tree. This architecture is defined by several key mechanisms:
Soft and Randomized Decision Routing
Unlike traditional decision trees that use hard "if-then" branches, R(Det)² utilizes soft routing. The model calculates routing probabilities, and the final prediction is a weighted sum of outputs from multiple nodes. This ensures the entire structure remains differentiable and trainable via standard backpropagation. To prevent the tree from collapsing into a state where all nodes learn identical features, the authors introduce Randomized Decision Routing, which injects stochasticity into the training process to encourage node diversity.
Dual Loss Optimization
To guide the learning of the decision tree, two specialized loss functions are employed:
- Selective Loss: This identifies a "dominant" node during training and assigns randomized weights to the losses of different nodes. This creates a "slow-fast" learning dynamic, forcing nodes to diverge and capture distinct types of information.
- Associative Loss: This ensures that the final fused output of the entire tree is accurate, effectively guiding the optimization of the routing probabilities themselves.
Architectural Variants
The authors propose three primary variants of the R(Det)² head:
- R(Det)²-B (Basic): Adds narrow branches to generate routing probabilities.
- R(Det)²-M (Masked): Utilizes "routing masks" to provide different nodes with filtered versions of input features, further increasing feature diversity.
- R(Det)²-T (Task-aware): Combines the tree structure with task-disentanglement techniques to optimize classification and localization separately.
Experimental Results
Performance on MS-COCO
The R(Det)² framework was evaluated exclusively on the MS-COCO 2017 dataset. The results demonstrate that R(Det)² is a highly effective "plug-and-play" module compatible with various detector types:
- Faster R-CNN (ResNet-50): AP increased from 37.4% to 41.0% (+3.6%).
- Libra R-CNN: +3.1% AP improvement.
- Cascade R-CNN: +2.2% AP improvement.
- RetinaNet: +1.8% AP improvement.
When integrated with a Swin-Large backbone and Deformable Convolutions (DCN), the model achieved a peak performance of 57.4% AP on the COCO test-dev benchmark.
Key Insights
One of the most notable findings is that R(Det)² provides the most significant improvements for large objects ($AP_L$). This suggests that the multi-node approach is particularly adept at integrating high-resolution visual evidence that single-node heads might overlook. Additionally, the authors developed a "Lite" version of the head that provides a 2.8% AP boost with negligible increases in computational complexity (FLOPs).
Discussion and Limitations
While R(Det)² offers substantial accuracy gains, it does introduce additional parameters. The Task-aware version, for instance, increases parameters from approximately 15.62M to 45.97M. Although the "Lite" version mitigates this, there remains a fundamental trade-off between model complexity and detection precision. Furthermore, the model's performance is somewhat sensitive to the balance coefficient ($\lambda$) between selective and associative losses, typically requiring tuning around an optimal value of 0.5.
Conclusion
R(Det)² represents a significant shift in how detection heads are designed. By moving away from simple linear stacks toward randomized, hierarchical decision trees, the model achieves greater feature diversity and improved robustness. Future work may explore deeper tree topologies and the application of randomized routing to other computer vision tasks beyond object detection.
Summary updated