Segmentacija in rekonstrukcija objektov iz oblaka točk z uporabo globokih nevronskih mrež

Slabanja, Jurij (author), Solina, Franc (mentor)

Na področju računalniškega vida se je že zgodaj pojavila potreba po modeliranju vizualnih informacij s kompaktnimi modeli. Z razširitvijo vse bolj zmogljivih senzorjev za zajem vizualnih informacij je postalo to še posebej pomembno. V zadnjih nekaj letih je za hitro in učinkovito procesiranje takih informacij postala precej popularna uporaba nevronskih mrež. V tem delu smo implementirali konvolucijsko nevronsko mrežo, s katero lahko določimo ali vsaj aproksimiramo vse objekte v oblaku točk. Začeli smo s preprosto arhitekturo za napoved parametrov enega objekta. Nato smo mrežo razširili na arhitekturo podobno Faster R-CNN, s katero lahko napovemo parametre poljubno mnogo objektov v sceni. Objekte smo modelirali s superkvadriki. Rezultati za prvotno mrežo izgledajo precej obetavni. Za posplošeno mrežo so še vedno pretežno dobri, so pa razumljivo nekoliko slabši od prvotne mreže, saj kompleksnost problema naraste. Potrebno je segmentirati vse objekte med sabo, ne samo napovedati parametre za vsak posamezen objekt.

• Language: Slovenian
• Keywords: racunalniški vid, segmentacija, 3D rekonstrukcija, oblak tock, globoke nevronske mreže, TensorFlow, Keras
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2017
• PID: 20.500.12556/RUL-96727
• Publication date in RUL: 12.10.2017

Secondary language

• Language: English
• Title: Segmentation and reconstruction of objects from a point cloud with deep neural networks

Abstract

The need to model visual information with compact representations has existed since the early days of computer vision. With the spread of ever more powerful and capable sensors, this need has become more and more present in recent years. As such, neural networks have become the popular choice for quick and effective processing of visual data. For this thesis we implemented a convolutional neural network with which we can determine or at least approximate all objects in a given point cloud scene. We started off with a simple architecture that could predict the parameters of a single object in a scene. Then we expanded it with an architecture similar to Faster R-CNN, that could predict the parameters for any amount of object in a scene. The results for the initial neural network were satisfactory. The second, generalized one still gave decent results, but compared to the initial one understandably performed somewhat worse, since it was also necessary to segment all the objects apart, not just predict parameters for each one.

• Keywords: computer vision, segmentation, 3D reconstruction, point cloud, deep neural networks, TensorFlow, Keras