ŠOOŠ, M. 3D point cloud segmentation for industrial bin-picking [online]. Brno: Vysoké učení technické v Brně. Fakulta strojního inženýrství. 2023.

Posudek vedoucího

Shehadeh, Mhd Ali

The master's thesis on 3D point cloud segmentation for industrial bin-picking by Ing. Marek Šooš successfully accomplishes the defined objectives, which include reviewing robotic 3D vision approaches for bin picking, developing a 3D point cloud instance segmentation application for the bin-picking scene, and performing registration of the segmented objects to estimate their pose. The thesis demonstrates the student's effective ability to build and train neural networks, as evidenced by the implementation of segmentation and registration techniques to determine the position of 3D CAD models. Notably, the student's innovative solution of generating a dataset specifically tailored for training purposes is commendable, as it eliminates the need for time-consuming manual sample collection and labelling. However, it is important to acknowledge that the generated dataset, although valuable, may introduce pose variations that are not typically encountered in real-life scenarios. Additionally, some refinements are required for the implemented overlapping avoidance technique to enhance its effectiveness. Throughout the year, the student has exhibited dedication and commitment, achieving the objectives with minor supervision. Given more time for collaborative review and discussion, further improvements could have been made, particularly in terms of enhancing readability and clarity. Regarding segmentation precision, there are some concerns raised by the presented results in Table 1, suggesting potential areas for strengthening the robustness of the segmentation algorithm. Nonetheless, I still consider the overall work to be of very good quality. In conclusion, Ing. Marek Šooš has successfully completed their master's thesis on 3D point cloud segmentation for industrial bin-picking. The thesis showcases his understanding of point cloud segmentation and proficiency in implementing segmentation and registration techniques. While minor language mistakes and areas for improvement were identified, the overall quality of the work is commendable. Consequently, I rate this thesis as being very good overall.

Posudek oponenta

Škrabánek, Pavel

The thesis presents an application of the fast point cloud clustering-based instance segmentation on various bin-picking tasks, and implementation of a dataset generator for training of the segmentation method. The thesis also contains a review of machine vision instrumentation used in bin-picking, and a brief review of segmentation techniques. From a formal point of view, the student has achieved objectives specified in the thesis assignment; however, the review of the segmentation techniques is shallow and does not contain the segmentation technique the student has implemented within the thesis. Note that all technical terms are correctly described (see for example subsubsection 2.5.5 – description of convolution), and the review of segmentation techniques is in fact based on one publication. The outline of the thesis is exemplary, but the content of the sections does not respect the outline. Specifically, a part of solution description can be found in literature review, and a great part of theory is in section 3, where the solution and experiments are presented. The text shows an effort to comply requirements on formal adjustment of the thesis, but not all requirements are met. For example, equations are not parts of sentences, functions are in italic, not all variables are introduced, variables in figures are not properly written, incorrect referencing of equations. The text is sometimes fluent but more often seems sketchy. The description of the solution is unfinished (e.g. the design of experiments and evaluation of experiments results). Probably because of that the evaluation of the results is rather based on expert (student) judgment than on numerical values. I find the ability of the student to make conclusions based on the results to be sub-standard (the discussion and the conclusion contain mostly a summary of the text outline). The appearance of the work is of a good standard. The work with citation is good. The text contains some grammatical mistakes, and the use of tenses does not respect scientific standards. The outputs of the work, especially the dataset generator, show an application potential. Considering all the facts, I recommend the thesis for the defence, and I evaluate it with the grade C.

eVSKP id 149277