KRSIČKA, T. Self-supervised přístupy pro efektivní analýzu 3D tvarů [online]. Brno: Vysoké učení technické v Brně. Fakulta informačních technologií. 2025.
Mr. Krsička achieved impressive results in his Master’s thesis, which holds the potential to contribute to the scientific community focused on 3D machine learning. He is a highly independent student, capable of firmly grasping complex scientific ideas and delivering insightful results.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Informace k zadání | The analysis of 3D surface models using neural networks remains challenging due to limited data availability. In the image domain, self-supervised approaches are commonly used to boost the performance of tasks such as classification and segmentation. However, literature on applying such methods to 3D geometry analysis is still very limited. The goal of Mr. Krsička’s thesis was to design and implement a self-supervised pretraining strategy using autoencoders and to evaluate its impact on a 3D shape classification task. This was an experimental and demanding assignment, especially considering that Mr. Krsička chose to validate the approach on a medical shape dataset. This first required significant effort to adapt the data to the new problem setting. Thanks to this choice, the work has the potential to contribute to improved clinical outcomes. | ||
| Práce s literaturou | The student expanded on the initial set of scientific articles by independently gathering and studying high-quality research papers related to the topic. | ||
| Aktivita během řešení, konzultace, komunikace | The student consulted regularly and frequently throughout both semesters. Particularly creditable was his engagement during the summer semester, during which Mr. Krsička pursued a research internship at Akita University in Japan. | ||
| Aktivita při dokončování | The work was completed ahead of schedule. My minor comments on the text were incorporated into the final version of the technical report. | ||
| Publikační činnost, ocenění | Further validation may lead to a conference publication. |
Tomáš Krsička was able to understand the potential problems of existing solutions in the field and propose his own concept of the pooling operator for graph autoencoders. He has the right direction of thinking, suggests reasonable approaches (such as the support balancing and the bottleneck redistribution for positional node coding) and can express the principles in a mathematical manner. He demonstrates his ability to handle research topics. It's unfortunate that the technical report gives the impression that there was no time left for review and fixing the issues.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Náročnost zadání | Although self-learning techniques in deep learning have been a significant trend in recent years, their application in 3D shape analysis, especially in connection with Graph Neural Networks, remains limited. For this reason, the assignment is more difficult. | ||
| Rozsah splnění požadavků zadání | According to the assignment's suggestion, the author focused on graph neural networks and graph autoencoders, drawing inspiration from GraphMAE (Hou et al., 2022). He designed and implemented his own pooling and unpooling operations, presenting their advantages in experiments. The author also conducted experiments with self-supervised pre-training of the encoder for the 3D mesh classification task, using an autoencoder trained on the 3D shape reconstruction task. This is an extension of the original assignment, pushing the tone of the work more into a scientific experiment. | ||
| Rozsah technické zprávy | |||
| Prezentační úroveň technické zprávy | 75 | The structure of the technical report is well-organised. Chapter 2 provides a concise overview of the principles of GNN, GCN, and GAT models, summarising the key differences between their concepts. However, some details are placed in inappropriate chapters: The overview of existing local pooling strategies in Section 3.4 should be included in Chapter 2. In Chapter 3, which describes the author's own work, the novel pooling mechanism should be described and compared with the existing methods. Chapter 4 does not describe only implementation details. The design and principles of the PS pooling operator (Section 4.3) are methodological thoughts and should be included in the part describing the concepts. The loss function is not a minor "implementation detail" but rather an important part of the experiments. By the way, what are the normal consistency loss, the edge loss and the Laplacian smoothing terms in the loss? The adopted diagrams in Chapter 2, modified for graphs, are nice. It is a pity that the author did not accompany the description of his work with similar diagrams, which would have helped in understanding the detailed principles of PS pooling (Section 4.3) as well as the architecture of the models he experimented with. All this is only in text. Some figures have incorrect descriptions (e.g., columns vs. rows in Figure 5.10). | |
| Formální úprava technické zprávy | 65 | The text contains many errors. Very often, there is a small letter "a" at the beginning of a new sentence. I came across artefacts of unfinished sentences or meaningless word order. One such artefact is the beginning of the extended abstract in Czech. These unnecessary errors, which a spell checker can easily correct, harm the otherwise good impression of the English text, which is fluent when describing a solution or an experiment. | |
| Práce s literaturou | 85 | The student explored scientific articles on graph networks recommended in the assignment and also studied more general principles of methods for 3D shape representation and self-supervised techniques. He thoroughly understood these methods and was able to effectively build upon them with his design of a pooling mechanism for graph autoencoders. He could have paid more attention to architectures that utilise transformers, such as MeshMAE, and looked for at least one state-of-the-art (SotA) point cloud-based autoencoder for comparison in experiments. | |
| Realizační výstup | 90 | The work has the character of a research experiment. The author experimented with various architectures of graph neural networks and demonstrated the potential of utilising self-learning to pre-train graph autoencoders for the task of mesh classification. The experiments are generally well designed, but not everything is adequately described. Did you use any data augmentation? How were the optimal hyperparameters identified and selected? The Python code itself is well-structured and commented. | |
| Využitelnost výsledků | The author proposed a novel pooling operator designed for 3D meshes, motivated by the disadvantages of existing mechanisms, and demonstrated promising results in experiments. It would be great to also see an ablation study that clearly shows the benefits of the author's sound design decisions, such as the bottleneck redistribution (Section 3.6). Another practical result and potential contribution to the scientific community is the MedShapeNet19 dataset created by the author, who selected shapes from the original MedShapeNet dataset, which contains many problematic meshes. If the author plans to make the dataset public, I would just recommend not decimating the models, maintaining the maximum possible quality. The reduction to 10k faces can be performed as a pre-processing step for concrete experiments. |
eVSKP id 161915