ZDRAVECKÝ, P. Generativní modely pro doplnění 3D tvaru [online]. Brno: Vysoké učení technické v Brně. Fakulta informačních technologií. 2024.
Mr. Zdravecký has dedicated considerable effort to tackling the issue of filling in missing parts of 3D models using generative models. This effort has resulted in a solution that extends the current state of knowledge in the field. Considering these factors alongside the complexity of the task itself, the outcomes of his work exceed the standards for a Master’s Thesis.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Informace k zadání | The topic of Mr. Zdravecký's Master’s Thesis is highly experimental. The interdisciplinary nature of applying generative deep learning concepts to 3D geometry makes this assignment particularly complex and non-trivial. Solving it required a thorough understanding of very recent research works, as well as substantial practical work since the foundational research he relied on does not provide any public implementation. The student enthusiastically engaged with this problem, demonstrating a deep understanding of the state-of-the-art concepts and their interconnections. He developed a high-quality codebase that facilitated easy experimentation with his solution, and ultimately, he presented his valuable results in a high-quality manner. | ||
| Práce s literaturou | The student was provided with literature forming the foundational research for his work. Understanding this material required an in-depth study of generative models and algorithms for processing 3D geometry. The student successfully sourced all necessary information from high-quality scientific articles published in reputable conferences and journals. He also placed his work within the context of other current approaches to using deep learning on 3D data by rigorously studying the current state of knowledge. | ||
| Aktivita během řešení, konzultace, komunikace | The student worked independently for the most part. Consultations with the supervisor were primarily for discussing scientific articles, not for explanation, but for critical reflection on the presented methods and results. Consultations also served to present the finished work and to confirm the next steps, which Mr. Zdravecký had thoughtfully planned in advance. | ||
| Aktivita při dokončování | The work was completed ahead of schedule, allowing me to review and comment on the entire technical work before it was submitted. The student incorporated these comments into the final version. | ||
| Publikační činnost, ocenění | The student follows the trend of reproducible research and publishes source files, prepared datasets and trained models as open-source. The student's work was accepted for the Excel@FIT student conference, where an expert panel recognized it with an award. |
Mr. Zdravecký shows that scientifically oriented work is natural for him, and he can handle it very well. Considering the difficult topic of a research nature, the quality of the work, and the results achieved, I consider the work excellent.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Náročnost zadání | The assignment is an active research topic. Studying and experimenting with state-of-the-art models for 3D shape completion is a challenging task and is not made easier by high demands on the computing power for training such models. | ||
| Rozsah splnění požadavků zadání | All required points have been met. The author followed the architecture of the DiffComplete model published in 2023 and, in addition to its implementation (the model is not available), he proposed two modifications: a user input (region of interest) for better guiding automatic shape completion and a multi-resolution approach for generating more detailed shapes. | ||
| Rozsah technické zprávy | The theoretical chapters, the proposed solution's description, and the discussion of the experiments are adequate. The author's effort to ensure the reproducibility of his experiments and results is evident. Only a few paragraphs could have been more brief, e.g., the description of TSDF generation from depth images (which is not essential for the work), the description of spatial 3D transformations (which is essential knowledge), or different surface smoothing techniques (the author uses Laplace anyway). | ||
| Prezentační úroveň technické zprávy | 80 | The author chose a clear structure for the technical report. The theoretical chapters and the overview of the state-of-the-art are very well prepared. Nice pictures suitably complement the whole text. Working with datasets and defining the experiments are not trivial, and because a lot of information is written in paragraphs of text, I sometimes miss the important ones. It would be helpful to further structure the text into bullet points to emphasize the most important information. At the end of the conducted experiments, I would appreciate the overall comparison of the different variants in one table; it would be more evident. | |
| Formální úprava technické zprávy | 85 | The technical report is written in very good English. I have no complaints about the work's typographical and linguistic aspects. Just some articles' citations are incomplete; the author probably found pre-prints available on arXiv and not the published versions. | |
| Práce s literaturou | 80 | The study literature is extensive, primarily scientific publications, and there are also the latest ones from recent years. The author mainly follows the DiffComplete diffusion model. Unfortunately, it does not seem he took much inspiration from other architectures, which is very nicely described in the overview of current approaches. | |
| Realizační výstup | 95 | The experiments carried out are of high quality and well-processed. The source codes of the scripts are largely modifications of previously published codes, which he followed up on and used very well. The whole project, including all source codes and models, is well described by the author and published on GitHub. | |
| Využitelnost výsledků | The diploma thesis has the character of a scientific work. The author followed up on a recently published method based on a diffusion model, which he extended in his experiments with a basic user annotation of the region of interest, which significantly helped to complete the correct part of the model and also with a multi-resolution approach, which allowed him to train a network capable of generating models in higher detail. The proposed procedures and experiments can serve to steer the research further. |
eVSKP id 154507