Introduction
Dense prediction tasks, such as semantic segmentation and depth estimation, are fundamental for autonomous systems but pose significant computational challenges for edge devices. While Multi-Task Learning (MTL) and Neural Architecture Search (NAS) have been explored independently to address efficiency, their combined potential remains largely untapped. This paper introduces EDNAS, a framework that synergistically optimizes model architecture for multiple tasks simultaneously while adhering to hardware constraints. By addressing the inherent instability in depth estimation training with a new loss function, JAReD, the authors achieve state-of-the-art performance with a fraction of the computational cost.
Methodology
The core contribution is a two-pronged approach focusing on architectural synergy and training stability.
EDNAS: Multi-Task Neural Architecture Search
Unlike traditional methods that adapt single-task models for multi-tasking, EDNAS searches for a backbone inherently designed for parameter sharing. The search space is built upon EfficientNet-B0 and BiFPN fusion modules, allowing the algorithm to select optimal combinations of standard and "fused" bottleneck layers. This "Searching for layers" approach is contrasted with traditional MTL paradigms in the figure below.
JAReD Loss for Training Stability
A significant hurdle in NAS for dense predictions is the noise in depth estimation metrics. To mitigate this, the authors proposed the Joint Absolute-Relative Depth (JAReD) loss. By incorporating a weighted relative-error component alongside standard absolute error, JAReD reduces evaluation noise by up to 88%, providing the stable signal necessary for the NAS controller to navigate a search space of approximately $1.2 \times 10^{24}$ architectures.
Results and Discussion
EDNAS demonstrates a clear Pareto improvement over existing methods. It achieves a relative accuracy gain of +8.5% to +10.9% on CityScapes while utilizing only 1/10th of the parameters and FLOPs compared to previous state-of-the-art models like AdaShare.
Key findings include:
- Synergistic Optimization: Models searched specifically for multi-tasking consistently outperform those optimized for single tasks and later adapted.
- Hardware-Awareness: The search favored "Fused-IBN" layers, which, despite higher parameter counts, offer superior latency on edge accelerators like the Coral TPU due to their optimization for dense convolutions.
- Robust Generalization: Performance gains were consistent across both indoor (NYU-v2) and outdoor (CityScapes) datasets.
Conclusion
EDNAS and JAReD provide a robust framework for deploying high-performance dense prediction models on resource-constrained edge hardware. Future work will explore the application of this synergy to mixed dense and sparse tasks and evaluate performance across a broader range of hardware accelerators.
Summary updated