Abstract
Accurate 3D road reconstruction is vital for autonomous navigation but is often hindered by the domain gap between generic training datasets and specific road environments. This paper presents D3Stereo, a plug-and-play framework designed to adapt pre-trained deep stereo networks to road scenarios without fine-tuning. By integrating Recursive Bilateral Filtering (RBF) for efficient cost aggregation and a novel Decisive Disparity Diffusion strategy, D3Stereo identifies and propagates highly reliable disparity "seeds" across scales. Experimental results on synthetic and real-world datasets demonstrate that D3Stereo significantly reduces error rates—by up to 83% in some metrics—while maintaining computational efficiency, outperforming specialized road-reconstruction baselines.
Introduction
Autonomous vehicles rely on precise 3D road surface data to detect hazards like potholes and cracks. While LiDAR offers high precision, its cost limits widespread adoption. Stereo vision provides a cheaper alternative but faces significant challenges: data scarcity for labeled road disparities, generalization failures of Deep Convolutional Neural Networks (DCNNs) trained on generic data, and the high computational cost of traditional explicit algorithms. D3Stereo addresses these by leveraging the feature-extraction power of existing DCNNs and adapting them to road textures through a robust propagation-based mechanism.
Methodology
The D3Stereo framework is built upon three core technical pillars designed for efficiency and reliability:
- Recursive Bilateral Filtering (RBF): To overcome the memory and time constraints of standard bilateral filtering, the authors utilize a recursive approach stacking $3 \times 3$ kernels. This mimics a large receptive field while drastically reducing GPU overhead.
- Intra-scale Decisive Disparity Diffusion: The system identifies "decisive" pixels with high-confidence disparities using Peak Ratio Naive (PKRN) scores. These seeds diffuse to neighbors via an adversarial mechanism that allows for the correction of previous estimates if more consistent data emerges.
- Inter-scale Inheritance: Reliable disparities are passed from coarse to fine layers in a feature pyramid, narrowing the search space for high-resolution matching and filtering out noise.
Experimental Results
D3Stereo was evaluated on the synthetic UDTIRI-Stereo dataset (featuring realistic road noise and potholes), the real-world Stereo-Road dataset, and the Middlebury benchmark. The results show that D3Stereo acts as a powerful performance booster for existing networks. For instance, when integrated with PSMNet or AANet, it reduced End-Point Error (EPE) by up to 63.10% on road data.
The radar charts in Figure 3 highlight that D3Stereo consistently improves the EPE and runtime efficiency across various backbones. Notably, even a simple VGG-16 backbone pre-trained on ImageNet, when used within the D3Stereo framework, can outperform specialized road-reconstruction algorithms like PT-SGM.
Conclusion
D3Stereo provides a robust, zero-shot solution for adapting deep stereo networks to the unique challenges of road environments. While the current diffusion process can lead to "edge-fattening" at object boundaries, future integration with semantic segmentation promises to resolve these artifacts. The framework's ability to significantly enhance pre-trained models without additional labeled data makes it a highly practical tool for the next generation of autonomous driving systems.
Summary updated