Hyperparameter optimization of an augmented autoencoder for 6D object pose estimation via neural architecture search

Deep learning has become a cornerstone in numerous applications, such as robotics, augmented reality, and autonomous systems, where accurate 6D pose estimation – determining an object's position and orientation in 3D space – is critical. Despite its success, designing optimal neural architectures and tuning hyperparameters remain computationally expensive and challenging, especially in high-variance and data-intensive domains like 6D pose estimation. To address this, we investigate the application of a Neural Architecture Search (NAS) technique guided by classical Machine Learning-based performance predictors. As these predictors estimate final model performance using early-stage training data, we demonstrate that leveraging such a strategy allows for efficient exploration of the hyperparameter space, significantly reducing the computational burden of exhaustive search while still achieving improved performance. Building on an existing NAS method, we introduce a novel modification that enhances the efficiency of the search process, further accelerating convergence by nearly $ 71% $ without sacrificing accuracy. Through extensive experimentation on the LineMOD dataset, we demonstrate that our method consistently discovers high-performing configurations of an Augmented Autoencoder for 6D pose estimation, outperforming benchmark models by almost $ 15% $ in pose accuracy and $ 42% $ in reconstruction loss. These results underscore the potential of predictor-based NAS as a powerful and computationally efficient tool for neural architecture optimization in complex, real-world tasks.