LEČBYCH, M. Computer Vision for Autonomous Vehicles [online]. Brno: Vysoké učení technické v Brně. Fakulta strojního inženýrství. 2022.

Posudek vedoucího

Shehadeh, Mhd Ali

The student fulfilled the objectives of the thesis "Computer Vision for Autonomous Vehicles”. by programming his own CNN model and training it on PASCAL VOC dataset. His work is more than satisfactory, since the developed model outperforms YOLOv3 with better mean average precision (mAP). Moreover, the added features, namely lane detection and distance estimation, constitute a core for a useful autonomous driving system. However, the model needs to be adjusted for various illumination conditions (it is a popular challenge in the field of computer vision). Unfortunately, the student postponed most of the work to the last month, therefore, he could not allocate enough time for that. The student completed the work independently with minimal instructions from the thesis supervisor. Graphically, stylistically, and spelling-wise, the thesis is of satisfactory standard. In conclusion, I would like to highlight the graduate's ability to independently solve the assigned task. I recommend the thesis for defence with grade B.

Posudek oponenta

Škrabánek, Pavel

The thesis presents an object detector, a lane detector, and a distance estimator. All three computer vision systems are intended for autonomous vehicles. In the theoretical part of the thesis, the student summarizes computer vision tasks to be handled in autonomous vehicles. Thus, the student achieved all the objectives specified in the thesis assignment. The presented object detector is a modified version of well-known YOLO object detector. Both detectors are compared on an identical dataset. The thesis contains a survey on object detectors; however, the survey does not include the state-of-the art solutions. The presented distance estimator is not compared with any existing solution, and no survey on distance estimation is provided. The same is true for the lane detector. The evaluation results are not discussed, and no conclusion is based on the discussion. This is probably due to the vaguely specified goal of the work. Specifically, the reader will not know with the text what and for what purpose it should have been created. This fact reflects in the structure of the text which is not appropriately designed. Specifically, the survey on autonomous vehicles is mixed with a deep learning theory in one section. The rest of theory blends with solutions in remaining sections. Also purpose of section 6 “perception system” will be a mystery for the readers. The provided theory is shallow with minimum math which is, however, essential for this topic. The variables and equations do not respect the norm. The text formation is not consistent (see first lines in seconds and subsequent paragraphs within all sections and subsections). Abbreviations used in the text are not properly defined. The student works with citations adequately; however, some citations are missing (citation of used datasets). Thanks to the missing comparison of the proposed solutions with existing ones, it is hard to evaluate contribution and originality of the student. The only exception is the object detector where the student clearly shows his contribution. The modified object detector has an application potential. Considering all the facts, I recommend the thesis for the defence, and I evaluate it with the grade D.

eVSKP id 140195