Get started
Work

Estimation de la distribution de la pose d'un objet à 6 degrés de liberté à partir d'une seule image RGB et du modèle 3D de l'objet

Year: 2025
Type: dissertation
Abstract: Estimating the six degrees of freedom (6DoF) pose of object is a crucial task for a wide range of applications, including robotic grasping and manipulation, augmented reality, and autonomous navigation. While existing methods have achieved remarkable results, most state-of-the-art approaches are restricted to predicting a single pose. This formulation is ill-suited for real-world scenarios where symmetries, occlusions, sensor noise, or even poor lighting introduce visual ambiguity, allowing multiple plausible poses to explain the observed image. Furthermore, although generalization to unseen objects has been widely studied in the context of single-pose estimation, the extension of pose-distribution methods to unseen objects has, to the best of our knowledge, not yet been explored. Another limitation lies in current benchmarks and evaluation metrics, which typically assume that object symmetries are independent of the image. In practice, observable symmetries depend on image-specific factors such as viewpoint, occlusion, and image resolution. Ignoring these dependencies can lead to inconsistent or misleading evaluations. Finally, there is no unified framework for evaluating pose-distribution methods that enables fair comparison of different approaches under common metrics and reliable pose-distribution ground truths.This thesis aims to address these challenges by focusing on the estimation of 6D pose distributions of rigid objects from a single RGB image, assuming that accurate 3D object models are available.The first contribution is BOP-Distrib, a novel framework for generating per-image 6D pose distribution ground truth and adapting evaluation metrics accordingly. Unlike existing benchmarks, BOP-Distrib produces image-specific annotations that account for symmetries and occlusions, resulting in more realistic and reliable ground truth. A new metric formulation is also proposed to enable the evaluation of pose-distribution methods. Thus, BOP-Distrib establishes the first benchmark for pose-distribution estimation on real images and shows that re-evaluating state-of-the-art single-pose methods under this framework significantly changes their rankings.The second contribution is Corr2Distrib, the first indirect approach for 6D pose distribution estimation, which achieves state-of-the-art performance. This method assigns both a descriptor and a local frame to each 3D surface point of the object. Descriptors and local frames are trained to be symmetry-aware, allowing ambiguous 2D-3D correspondences, traditionally a limitation for PnP-based methods, to be directly exploited. Each correspondence generates a valid rotation hypothesis, and by grouping consistent correspondences that yield identical rotations, PnP can be applied efficiently to each group. In this way, Corr2Distrib reconstructs the full set of poses forming the distribution without requiring combinatorial search.The final contribution is ZeroDistrib, an extension of Corr2Distrib designed for unseen objects. In contrast to Corr2Distrib, where descriptors and local frames are object-specific, ZeroDistrib leverages foundational models such as DINOv2 to learn zero-shot features that generalize beyond the training set. At inference, given the CAD model of a novel object, a lightweight onboarding stage transfers these representations onto its 3D surface, enabling pose distribution estimation from 2D-3D correspondences without object-specific retraining. Preliminary experiments are presented to validate some components of the method and demonstrate its feasibility. A full implementation and large-scale evaluation of the framework are left as future work beyond the scope of this thesis.
Author: Brazi, Asma
Language: English
Open Access status: green
OpenAlex ID: W7143260937
Cites:
Cited by:
Related to: