JC
J.R.C. Campolatarro
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
2 records found
1
Interactive Class-Specific Segmentation Refinement on Irregular 3D Data
A Fast Feed-Forward Mask-Guided Approach
Irregular 3D data representations, such as point clouds and surface meshes, are fundamental to various high-stakes applications that require accurate semantic segmentation. Although fully automatic models can achieve high performance, they often leave localized errors that require human correction. Existing interactive methods for irregular 3D data are primarily designed to segment objects from scratch without leveraging a pre-existing mask, while the only known framework directly addressing this relies on computationally expensive test-time optimization.
To address this research gap, this thesis proposes a fast, feed-forward Graph Neural Network for mask-guided, class-specific interactive segmentation refinement on irregular 3D data. Our EdgeConv-based architecture uses early fusion to concatenate geometric features with the previous or external binary segmentation mask and dense distance maps encoding user interactions. This allows corrective user input to be processed through a single forward pass without updating the model parameters during inference. To study robustness, we introduce a multi-tier initialization strategy and a decoupled click-simulation framework that independently controls click allocation and placement.
We evaluated the method primarily on IntrA medical surface meshes for aneurysm segmentation and additionally on two binary part-segmentation subsets of ShapeNetPart. Across datasets and initialization qualities, it improved input masks and reached an average foreground IoU of approximately 90% or higher within 20 interactions.
Cross-evaluation of training and testing simulations shows that training on structured interaction heuristics provides no advantage over purely random clicks or an exact mirror of the standard testing protocol. Under the standard benchmark, random-click training matches the mirrored protocol while being more robust to alternative stochastic interaction policies. On IntrA, the model further improves external EPT-Net predictions from 88.82% to 99.52% foreground IoU within 20 interactions. Accurate initial masks reduce worst-case interaction effort relative to a matched segment-from-scratch model, whereas severely flawed masks can be less efficient than beginning from an empty state. However, all evaluated workflows converge to similar final accuracy, and the refinement model also handles empty masks as a natural fallback.
These results demonstrate that class-specific, feed-forward interactive refinement is feasible across irregular 3D representations and offers an efficient bridge between fully automatic segmentation and interactive segmentation from scratch.
Irregular 3D data representations, such as point clouds and surface meshes, are fundamental to various high-stakes applications that require accurate semantic segmentation. Although fully automatic models can achieve high performance, they often leave localized errors that require human correction. Existing interactive methods for irregular 3D data are primarily designed to segment objects from scratch without leveraging a pre-existing mask, while the only known framework directly addressing this relies on computationally expensive test-time optimization.
To address this research gap, this thesis proposes a fast, feed-forward Graph Neural Network for mask-guided, class-specific interactive segmentation refinement on irregular 3D data. Our EdgeConv-based architecture uses early fusion to concatenate geometric features with the previous or external binary segmentation mask and dense distance maps encoding user interactions. This allows corrective user input to be processed through a single forward pass without updating the model parameters during inference. To study robustness, we introduce a multi-tier initialization strategy and a decoupled click-simulation framework that independently controls click allocation and placement.
We evaluated the method primarily on IntrA medical surface meshes for aneurysm segmentation and additionally on two binary part-segmentation subsets of ShapeNetPart. Across datasets and initialization qualities, it improved input masks and reached an average foreground IoU of approximately 90% or higher within 20 interactions.
Cross-evaluation of training and testing simulations shows that training on structured interaction heuristics provides no advantage over purely random clicks or an exact mirror of the standard testing protocol. Under the standard benchmark, random-click training matches the mirrored protocol while being more robust to alternative stochastic interaction policies. On IntrA, the model further improves external EPT-Net predictions from 88.82% to 99.52% foreground IoU within 20 interactions. Accurate initial masks reduce worst-case interaction effort relative to a matched segment-from-scratch model, whereas severely flawed masks can be less efficient than beginning from an empty state. However, all evaluated workflows converge to similar final accuracy, and the refinement model also handles empty masks as a natural fallback.
These results demonstrate that class-specific, feed-forward interactive refinement is feasible across irregular 3D representations and offers an efficient bridge between fully automatic segmentation and interactive segmentation from scratch.
Mesh data is widely used in engineering for instance for simulations, CAD engineering and visualizations. The accuracy and quality of the meshes influence the reliability and validity of these processes. Besides manual modelling, scanning is becoming increasingly more common due to the increase in devices that have scanning capabilities. Unwanted noise is often present in scanned models. The process of mesh denoising is removing the unwanted noise whilst keeping the features of the mesh. These features are often anisotropic, e.g. sharp edges and corners.
The DeltaConv convolution is an anisotropic convolution, Wiersma et al. [24] show the advantage of using the anisotropic DeltaConv convolution for anisotropic tasks over other isotropic convolutions. In this thesis it is investigated if state-of-the-art mesh denoising can benefit from using the DeltaConv convolution. This is done by integrating the DeltaConv convolution in the Dual-DMP [6] algorithm, and tuning this network.
In this thesis we found that state-of-the-art mesh denoising can benefit from using the DeltaConv convolution. Due to the expressiveness of the DeltaConv convolution, objects with sharp features are denoised better than the state-of-the-art algorithms and on smooth meshes, DeltaConv works comparable to state-of-the-art algorithms. ...
The DeltaConv convolution is an anisotropic convolution, Wiersma et al. [24] show the advantage of using the anisotropic DeltaConv convolution for anisotropic tasks over other isotropic convolutions. In this thesis it is investigated if state-of-the-art mesh denoising can benefit from using the DeltaConv convolution. This is done by integrating the DeltaConv convolution in the Dual-DMP [6] algorithm, and tuning this network.
In this thesis we found that state-of-the-art mesh denoising can benefit from using the DeltaConv convolution. Due to the expressiveness of the DeltaConv convolution, objects with sharp features are denoised better than the state-of-the-art algorithms and on smooth meshes, DeltaConv works comparable to state-of-the-art algorithms. ...
Mesh data is widely used in engineering for instance for simulations, CAD engineering and visualizations. The accuracy and quality of the meshes influence the reliability and validity of these processes. Besides manual modelling, scanning is becoming increasingly more common due to the increase in devices that have scanning capabilities. Unwanted noise is often present in scanned models. The process of mesh denoising is removing the unwanted noise whilst keeping the features of the mesh. These features are often anisotropic, e.g. sharp edges and corners.
The DeltaConv convolution is an anisotropic convolution, Wiersma et al. [24] show the advantage of using the anisotropic DeltaConv convolution for anisotropic tasks over other isotropic convolutions. In this thesis it is investigated if state-of-the-art mesh denoising can benefit from using the DeltaConv convolution. This is done by integrating the DeltaConv convolution in the Dual-DMP [6] algorithm, and tuning this network.
In this thesis we found that state-of-the-art mesh denoising can benefit from using the DeltaConv convolution. Due to the expressiveness of the DeltaConv convolution, objects with sharp features are denoised better than the state-of-the-art algorithms and on smooth meshes, DeltaConv works comparable to state-of-the-art algorithms.
The DeltaConv convolution is an anisotropic convolution, Wiersma et al. [24] show the advantage of using the anisotropic DeltaConv convolution for anisotropic tasks over other isotropic convolutions. In this thesis it is investigated if state-of-the-art mesh denoising can benefit from using the DeltaConv convolution. This is done by integrating the DeltaConv convolution in the Dual-DMP [6] algorithm, and tuning this network.
In this thesis we found that state-of-the-art mesh denoising can benefit from using the DeltaConv convolution. Due to the expressiveness of the DeltaConv convolution, objects with sharp features are denoised better than the state-of-the-art algorithms and on smooth meshes, DeltaConv works comparable to state-of-the-art algorithms.