Študent si samostatne vyhľadal relevantné zdroje a moderné metódy týkajúce sa danej problematiky. Dokázal úspešne naštudovať odbornú literatúru, rýchlo sa zorientovať v riešenom probléme a získané poznatky následne aplikovať vo svojej práci.

Práca bola vypracovaná postupne počas celého roka a bola dokončená v stanovenom termíne s prekonzultovaným obsahom.

The technical report presented by the student is easy to read, and the author communicates the goals and results of the thesis clearly.

The reader is first introduced to classical surface reconstruction methods, as well as to common artifacts that may occur in input point clouds. The author then presents several approaches to machine-learning-based surface reconstruction. The selection of publications is relevant, and their description is handled well. I consider Section 2.4, which introduces the basics of neural networks, to be somewhat less necessary. If such a section is included, it should either provide a more systematic overview or be more clearly focused on concepts directly needed for the thesis. In its current form, it discusses only selected aspects of neural networks, such as activation functions, convolutional networks, and encoder-decoder architectures, without making fully clear why exactly these topics were chosen and placed at the same level of importance.

The selected method is described in a dedicated chapter, which I find appropriate. The explanation is sufficiently detailed and easy to follow.

The implementation and experimental details are communicated well. The results are presented using clear 3D visualizations, as well as suitable graphs and tables where appropriate.

The work resulted in two main outputs. The first is a set of Python scripts implementing an optimized version of the I-Filtering framework for surface reconstruction from point clouds, including data preparation, training, and evaluation. The second main output is the set of experimental results presented in the technical report, including verification against the results of the original solution as well as additional custom experiments. The experiment studying sensitivity to model capacity would be more informative if it were more clearly separated whether the observed behaviour is caused by the complexity of the model itself or by the different number of sampled points across datasets.

Evaluation level: zadání splněno

The assignment has been fulfilled. The student provided an overview of suitable deep learning methods for the target task and selected a relevant recent approach. The chosen method was implemented by the student, optimized, and rigorously evaluated using both the original experiments and additional experiments defined by the student.

The work contains a good number of high-quality references. Some references are cited as arXiv preprints, even though peer-reviewed versions are available. In addition, some web sources are not cited completely, as they do not include access dates.