| Meno: | Ján |
|---|---|
| Priezvisko: | Kamas |
| Názov: | 3D Object Pose Estimation in Structured Point Clouds with Uncertainty Quantification |
| Vedúci: | Ing. Viktor Kocur, PhD. |
| Rok: | 2026 |
| Kµúčové slová: | 3D object pose estimation, uncertainty quantification, point clouds, rotation |
| Abstrakt: | In this work, we integrate uncertainty quantification methods for both translation and rotation into a neural network for 3D object pose estimation from structured point cloud data. We investigate multiple approaches for modeling epistemic uncertainty, including Monte Carlo dropout, Bayesian neural networks, and deep ensembles. In addition, we model aleatoric uncertainty by directly predicting dispersion parameters within the network. A practical contribution of this work is the adaptation of uncertainty estimation methods to handle rotation, which lies on the non-Euclidean manifold SO(3) and cannot be directly modeled using standard Gaussian assumptions. The considered methods are evaluated using standard uncertainty-aware metrics, including negative log-likelihood (NLL), continuous ranked probability score (CRPS), prediction interval coverage, and sharpness. Experiments were carried out on two datasets: a smaller dataset for the main experiments and a larger dataset for additional evaluation. The results highlight differences in epistemic uncertainty between the smaller and larger datasets, with uncertainty decreasing as the amount of training data increases, and show that aleatoric uncertainty increases in cases of ambiguous rotation (in our case, related to symmetrical objects captured within data). From a practical perspective, Monte Carlo Dropout represents a simple uncertainty estimation method to integrate into an existing model, while deep ensembles achieved the best overall uncertainty estimation performance at the cost of significantly higher computational requirements. Bayesian neural networks provided a good compromise between uncertainty quality and efficiency, particularly when fine-tuned from a pretrained deterministic model. |
Súbory diplomovej práce:
| masters-thesis.pdf |
Súbory prezentácie na obhajobe: