Abstract
We introduce FPG-NAS, a FLOPs-aware Gated Differentiable Neural Architecture Search framework for efficient 6DoF object pose estimation. Estimating 3D rotation and translation from a single image has been widely investigated yet remains computationally demanding, limiting applicability in resource-constrained scenarios. FPG-NAS addresses this by proposing a specialized differentiable NAS approach for 6DoF pose estimation, featuring a task-specific search space and a differentiable gating mechanism that enables discrete multi-candidate operator selection, thus improving architectural diversity. Additionally, a FLOPs regularization term ensures a balanced trade-off between accuracy and efficiency. The framework explores a vast search space of approximately 10<sup>92</sup> possible architectures. Experiments on the LINEMOD and SPEED+ datasets demonstrate that FPG-NAS-derived models outperform previous methods under strict FLOPs constraints. To the best of our knowledge, FPG-NAS is the first differentiable NAS framework specifically designed for 6DoF object pose estimation.
Introduction
Estimating an object's 6DoF pose—its exact 3D position and orientation—is a cornerstone of robotics, augmented reality, and space navigation. However, the most accurate models are often computationally expensive, requiring massive amounts of memory and processing power (measured in FLOPs). This makes them difficult to deploy on resource-constrained hardware like drones or satellites.
While Neural Architecture Search (NAS) has successfully automated the design of efficient models for image classification, it hasn't been widely applied to 6DoF pose estimation. Standard NAS methods also suffer from a "winner-takes-all" flaw: during the design process, they force the system to choose only one type of operation per layer. FPG-NAS addresses this by allowing multiple neural blocks to work together in a single layer, discovering architectures that are significantly faster and lighter than state-of-the-art models without sacrificing accuracy.
Methods: Gated Differentiable Search
The authors introduce FPG-NAS (FLOPs-aware Gated Differentiable NAS), which features several key innovations:
- Differentiable Gating: Instead of using a standard "softmax" approach that picks one winner, FPG-NAS uses "Polarized Gates." These gates allow the search algorithm to select a combination of synergistic blocks within a single layer. This allows the model to choose multiple specialized operations that together outperform one large, generic operation.
- FLOPs-Aware Loss Function: The researchers added a specific penalty to the training process. If the architecture starts exceeding a pre-defined FLOPs budget, the loss function penalizes it, forcing the search to find a more efficient path.
- Task-Specific Search Space: The framework explores a massive space of $10^{92}$ possible architectures, specifically looking for the best "backbone" (feature extractor) and the best "fusion layer" (how the model combines features from different scales), which is critical for precise keypoint detection.
Macro-Architecture and Micro-Blocks
The individual blocks within the search space are inspired by EfficientNet and GhostNet. They utilize "Ghost Modules" to expand channels cheaply and depthwise convolutions to keep the computational load light. The search process optimizes both the backbone stages and the fusion mechanism (Simple Fusion vs. SE Fusion) to ensure high spatial precision for 3D keypoint localization.
Experimental Results
The framework was evaluated on two major benchmarks: LINEMOD (household objects in clutter) and SPEED+ (spacecraft pose estimation). The models discovered by FPG-NAS (named OursA and OursB) set new efficiency standards:
- LINEMOD: "OursA" achieved an average accuracy of 90.69% using only 10.03 GFLOPs. For comparison, the popular PVNet model requires 72.7 GFLOPs (over 7x more computation) to reach only 86.27% accuracy.
- SPEED+: FPG-NAS produced the most efficient model in the comparison (10.0G FLOPs) while maintaining the best overall accuracy on the Synthetic and Lightbox subsets.
- Versus Other NAS: FPG-NAS consistently outperformed standard NAS baselines (like DARTS) under the same computational budgets, proving that the "gated" approach is superior for this specific task.
Conclusion
FPG-NAS demonstrates that allowing a layer to use multiple synergistic blocks leads to better performance than forcing a single choice. By treating the computational budget as a hard constraint within the optimization loop, the system discovers "sweet spots" in architecture design that human engineers might overlook. Future work will explore block-merging strategies to further simplify discovered models and testing on diverse hardware targets like mobile chips or FPGAs.
Summary updated