APES-SOFT: uma abordagem eficiente em classificação de objetos 3D em nuvens de pontos com redes neurais convolucionais

• Ano de defesa: 2024
• Autor(a) principal: Silva, Iágson Carlos Lima
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Redes neurais (Computação); Nuvens de ponto; Sistemas de computação; Testes de desempenho; Neural Networks (Computing); Point clouds; Computing systems; Performance tests
• Link de acesso: http://repositorio.ufc.br/handle/riufc/78667

This work proposes an efficient method for the classification of 3D objects in point clouds using CNNs. The main objective is to propose a neural network called APES-Soft to enhance computational performance and efficiency in classifying 3D objects in point clouds based on the ablation analysis of the APES network. The adopted approach involves identifying the most relevant features for classification through ablation analysis. For this purpose, three distinct scenarios were conducted (Scenario I, Scenario II, and Scenario III), each adjusting diferente aspects of the network architecture to define the best candidate to be proposed as APES-Soft. The scenarios were conducted with rigorous evaluation, including statistical tests and traditional classification metrics. The results revealed that Scenario II stands out as an effective choice for the classification of 3D objects in point clouds, achieving higher accuracy among the scenarios, reaching 93.8% and presenting superior metrics compared to Scenarios I and III, APES (Global), and APES (Local), achieving a precision of 93.7%, the sensitivity of 93.8%, F1-Score of 93.7%, MCC of 93.5%, and Jaccard Index of 89.2%, besides being equivalent to the best-related works such as CurveNet and DeltaConv. Additionally, regarding computational performance, Scenario II showed a significant reduction in time compared to almost all scenarios, requiring only 20.35 hours of training while obtaining a memory consumption reduction of 21.89% during model training. Statistical tests, including ANOVA, Tukey’s HSD, Kruskal-Wallis Test, and Friedman Test, were performed to validate the model and compare the differences between the proposed methods. The results of these tests indicated that the observed differences between the methods were not statistically significant, suggesting a statistical equivalence among them. This study contributes to advancing 3D object classification and offers valuable insights into the essential features for accurate classification in point clouds. Moreover, it provides a deeper understanding of the balance between computational efficiency and model performance, a crucial aspect for practical viability in various applications