Attention-Augmented PointNet with Hybrid Ant Colony Optimization and Bayesian Hyperparameter Search for Robust 3D Point Cloud Classification in Augmented Reality Applications

Three-dimensional point cloud classification is a foundational task for augmented reality (AR), autonomous navigation, and robotics, yet remains challenging due to the inherent irregularity, sparsity, and permutation invariance of point set data. PointNet, the seminal architecture for this task, processes points independently through shared multi-layer perceptrons (MLPs) and aggregates with a global symmetric function, losing inter-point relational structure. We propose AA-PointNet-HBO, a framework that simultaneously addresses two under-explored aspects: (1) architectural expressiveness, by inserting a multi-head self-attention (MHSA) block between the per-point MLP encoder and the global aggregator so that each latent point embedding can dynamically attend to every other point, capturing long-range semantic co-occurrences; and (2) hyperparameter optimization, by introducing a two-tier Hybrid ACO–Bayesian Optimization (HBO) strategy that uses Ant Colony Optimization (ACO) to explore the discrete architectural design space (attention heads, channel widths, pooling strategy) and subsequently applies Bayesian Optimization (BO) with a Gaussian Process (GP) surrogate and Expected Improvement (EI) acquisition to refine the continuous training hyperparameter subspace (learning rate, dropout probability, weight decay). Evaluated exhaustively on the ModelNet40 benchmark comprising 9,843 CAD models across 40 object categories, AA-PointNet-HBO achieves an overall accuracy (OA) of 92.4% (mean ± std: 92.4 ± 0.18% over five independent seeds) and a mean class accuracy (mAcc) of 89.7%, surpassing the PointNet baseline by 3.3 percentage points with only 34%additional parameters (4.7M vs. 3.5M). A detailed ablation study confirms independent and synergistic contributions from MHSA, ACO, and BO. Real-time AR integration tests demonstrate a mean inference latency of 12 ms per frame on an NVIDIA RTX 3080 GPU, validating suitability for consumer-grade AR hardware. Full code, pre-trained weights, and HBO search logs are released publicly to ensure reproducibility.

@article{Almaghthawi2026,
  title     = {Attention-Augmented PointNet with Hybrid Ant Colony Optimization and Bayesian Hyperparameter Search for Robust 3D Point Cloud Classification in Augmented Reality Applications},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Ahmed Almaghthawi and Fahad Alharbi and Radwan M. Batyha and Suhaila Abuowaida},
  doi       = {10.14569/IJACSA.2026.0170703},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170703}
}