The goal of the presented work is to make a classifier based on neural networks that can discriminate between various objects presented as 3D pointclouds. The model is based on the well established Res-Net architecture structure where the basic residual block is taken form the RandLA-net network that was conceived with the processing of pointclouds in mind. The data, in the form of lists of x-y-z points, is analyzed by the model with subsequently broader and broader scopes to extract local features as well as global ones of the 3D objects. The model is trained from the ground up with the optimization of some hyperparameters. The net achieves almost state of the art performance in the benchmark ModelNet10 dataset.
L’obiettivo di questo lavoro è la creazione di un classificatore basato sulle reti neurali che riesca a distinguere oggetti diversi sotto forma di nuvole di punti. Il modello è basato sulla rinomata architettura Res-Net dove il blocco residuo fondamentale è stato preso dalla rete RandLA-net, poiché pensato apposta per l'analisi delle nuvole di punti. I dati, nella forma di liste di coordinate 3D, sono esplorati dal modello con un campo visivo via via più ampio in modo da estrarre proprietà locali e globali dell'oggetto analizzato. Il modello è stato allenato da zero con l'affinamento di qualche hyperparametro. I risultati della rete sono comparabili con lo stato dell'arte utilizzando il ModelNet10 come test.
Classificazione di modelli 3D con Deep Learning
PICCOLI, MARCO ANNIBALE
2021/2022
Abstract
The goal of the presented work is to make a classifier based on neural networks that can discriminate between various objects presented as 3D pointclouds. The model is based on the well established Res-Net architecture structure where the basic residual block is taken form the RandLA-net network that was conceived with the processing of pointclouds in mind. The data, in the form of lists of x-y-z points, is analyzed by the model with subsequently broader and broader scopes to extract local features as well as global ones of the 3D objects. The model is trained from the ground up with the optimization of some hyperparameters. The net achieves almost state of the art performance in the benchmark ModelNet10 dataset.File | Dimensione | Formato | |
---|---|---|---|
Piccoli_Marco_Annibale.pdf
accesso aperto
Dimensione
1.96 MB
Formato
Adobe PDF
|
1.96 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/39021