Master's Thesis

3D Shape Completion using Deep Learning

Final Thesis 16.02 MB

Author of thesis: Ing. Tomáš Rajsigl

Acad. year: 2025/2026

Supervisor: doc. Ing. Michal Španěl, Ph.D.

Reviewer: Ing. Dávid Pukanec

Abstract:

Real-world 3D scans often suffer from missing geometry, necessitating automated shape completion. This work investigates the application of deep learning methods for reconstructing such partial inputs. However, processing 3D data is generally memory-intensive, which imposes a severe bottleneck on the achievable resolution. To address this, the proposed method introduces a novel approach operating on a sparse TSDF representation within a diffusion-based framework. The model is trained entirely in a self-supervised manner by dynamically simulating occlusions on complete 3D shapes. The approach is validated against a dense state-of-the-art baseline, DiffComplete. All of the proposed sparse variants match the baseline across IoU, Chamfer Distance, and F-Score metrics, while reducing the active voxel count by 36 to 54% depending on the chosen configuration. Ultimately, this efficiency enables doubling the spatial resolution to 64^3, capturing finer details under the same hardware budget.

Keywords:

3D computer vision, 3D shape completion, sparse volumetric representations, Signed Distance Function, sparse convolutions, generative deep learning, diffusion models, EDM, DiffComplete, ControlNet

Date of defence

23.06.2026

Result of the defence

Defended (thesis was successfully defended)

znamkaAznamka

Grading

A

Process of defence

Student nejprve prezentoval výsledky, kterých dosáhl v rámci své práce. Komise se poté seznámila s hodnocením vedoucího a posudkem oponenta práce. Student následně odpověděl na otázky oponenta a na další otázky přítomných. Komise se na základě posudku oponenta, hodnocení vedoucího, přednesené prezentace a odpovědí studenta na položené otázky rozhodla práci hodnotit stupněm A.

Topics for thesis defence

  1. In your work, you report parameter savings based on the percentage of active voxels at 32³ resolution (54.9%). How does this saving change at higher resolutions, such as 128³? Do you feel that this approach is scalable?
  2. What is the data distribution regarding the missing volume generated by your pipeline?
  3. In the evaluation of the different proposed architectures, you show that the Bidirectional architecture outperforms the Single-Branch architecture. Could this improvement be caused by the higher number of parameters in the Bidirectional model?

Language of thesis

English

Faculty

Department

Study programme

Information Technology and Artificial Intelligence (MITAI)

Specialization

Machine Learning (NMAL)

Composition of Committee

doc. Ing. Vítězslav Beran, Ph.D. (předseda)
prof. Ing. Hynek Heřmanský, Dr. Eng. (místopředseda)
doc. Ing. Ondřej Lengál, Ph.D. (člen)
doc. Ing. František Zbořil, Ph.D. (člen)
doc. Ing. Michal Bidlo, Ph.D. (člen)
RNDr. Marek Rychlý, Ph.D. (člen)

Supervisor’s report
doc. Ing. Michal Španěl, Ph.D.

It was a pleasure to supervise Mr Rajsigl's thesis. The consultations with him felt like discussions with a qualified researcher in the field. He also demonstrated passion for the topic and diligence in preparing the technical report. He has achieved interesting results with an indisputable personal contribution and the potential for publication at a scientific conference.

Evaluation criteria Verbal classification
Informace k zadání

In his work, Mr Rajsigl focused on 3D shape completion techniques using the so-called SDF representation and diffusion models. For training, he developed an effective pipeline for dynamically generating artificially damaged models. I consider the sparse SDF representation a key contribution of the work, which can reduce memory requirements during model training. This involved a complete rewrite of the original model, inspired by DiffComplete, using sparse operations (convolution, etc.). The assignment was fulfilled excellently without any reservations.

Aktivita při dokončování

Both the work and the technical report were completed ahead of time, and the final report was consulted sufficiently.

Publikační činnost, ocenění

The work was awarded by an expert panel at the Excel@FIT student conference.

Práce s literaturou

The student was very active in searching and studying literature. From our discussions, it was clear that he understood the field deeply, identified potential problems and issues with various solutions, and knew how to properly design his own model and experiments.

Aktivita během řešení, konzultace, komunikace

The student was proactive, frequently consulting and consistently bringing his own clear suggestions to consultations. It was always evident that he was working hard on a solution and enjoyed experimental work with a touch of research spirit.

Points proposed by supervisor: 100

Grade proposed by supervisor: A

Reviewer’s report
Ing. Dávid Pukanec

The student demonstrated the capability to work with scientific literature and possesses the knowledge to construct a method for shape completion. The student designed their own improved version of a diffusion model for the general shape completion task using SDF volumetric representation as its input. By implementing an augmentation pipeline that generates input on the fly, he trained the network in a self-supervised manner.


Moreover, the student was able to perform experimentation and critically evaluate the results to improve and explore additional possibilities for enhancing the proposed approach. Although some minor errors and imperfections are present, the final work is of high quality and provides useful insights into the problem of shape completion and into designing a method for the aforementioned task.

Evaluation criteria Verbal classification Points
Rozsah splnění požadavků zadání

Evaluation level: zadání splněno

Rozsah technické zprávy

Evaluation level: je v obvyklém rozmezí

Prezentační úroveň technické zprávy

The work is well organized, with all sections meaningfully placed, and the author correctly uses them to create a continuous and highly readable report. The work is accompanied by images, each serving a meaningful purpose and helping with understanding. It would be helpful to include a figure solely focused on displaying the various models that the pipeline can generate. Another imperfection is the grouping of different generative paradigms together with graph architectures and input modalities.

92
Formální úprava technické zprávy

The work is written very well and contains nearly no mistakes. Some of the images, particularly those showing qualitative evaluation, would be better in higher resolution.

95
Práce s literaturou

The author demonstrated the ability to work with and understand scientific literature. The work includes several key and well-established publications. However, some key methods in the field could have been explored further. Some statements in the work lack proper support, such as the claim that there is no need for a broader spatial context when using bounding boxes to define the region of interest, or that using a finite noise bank has a negligible impact on memorisation. In future, it would be better to support these claims.

87
Realizační výstup

The results of the work consist of experiments and ablation studies examining how the memory effectiveness of large diffusion models working with SDF can be improved. The author provides an ablation study for each decision and the corresponding impacts. One ablation is missing and would be insightful, the change from traditional DDIM and DDPM to EDM diffusion. It would be valuable to see the impact of this decision. Moreover, more thorough visualization and evaluation of the created dataset would be welcome, mainly the distribution of the amount of missing volume in the models. Furthermore, detailed parameter settings for the hole generation pipeline would be welcome. However, overall the work is done very professionally and has useful implications.

89
Využitelnost výsledků

The author showcased that their method can directly compete with DiffComplete in shape completion while using fewer parameters and thus consuming less GPU memory. The experiments and ablation study of the author's design decisions provide insightful information for designing new and improved completion diffusion networks. Mainly, the findings show that the two-branch diffusion architecture is inferior to the one-branch or the bidirectional dual-branch architecture. 

Náročnost zadání

Evaluation level: obtížnější zadání

In accordance with the stated objectives, the author conducted an in-depth investigation of shape completion and identified a fundamental challenge in processing 3D data using neural networks. The author addressed the issue of excessive memory consumption through multiple approaches, developing a model with fewer parameters while achieving improved accuracy. The experimental results are highly informative and provide valuable insights. The work was presented as a poster at Excel. Therefore, the author has fully completed all assignment requirements.

Topics for thesis defence:
  1. In your work, you report parameter savings based on the percentage of active voxels at 32³ resolution (54.9%). How does this saving change at higher resolutions, such as 128³? Do you feel that this approach is scalable?
  2. What is the data distribution regarding the missing volume generated by your pipeline?
  3. In the evaluation of the different proposed architectures, you show that the Bidirectional architecture outperforms the Single-Branch architecture. Could this improvement be caused by the higher number of parameters in the Bidirectional model?
Points proposed by reviewer: 93

Grade proposed by reviewer: A

Responsibility: Mgr. et Mgr. Hana Odstrčilová