請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78640
標題: | 應用於三維點雲分類任務之注意力機制及神經網路搜索架構 Attention Mechanism and Neural Architecture Search for Three-dimensional Point Cloud Classification |
作者: | Yen-Po Lin 林彥伯 |
指導教授: | 盧奕璋(Yi-Chang Lu) |
關鍵字: | 點雲分類,注意力機制,動態 K 值調整,神經網路搜索, Point Cloud Classification,Attention Mechanism,Dyanmic K,Neural Architecture Search, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 近幾年來,隨著對自動駕駛技術的投入,三維點雲的研究也隨之 蓬勃發展。其中由於 3D 點雲有著不規則以及無順序的特性,因此要 抓取點與點之間的幾何特徵是非常困難的。本論文提出了 3 種方法 來改善抓取點雲特徵的能力,進而提升點雲分類任務的正確及穩定 度。在第一個方法中我們引入了 2 種不同面向的注意力機制,分別為 用來決定點與點之間關聯性大小的點注意力模組 (Point-wise Attention Module) 以及讓模型在有限資源下更專注於重要特徵的通道注意力模 組 (Channel-wise Attention Module)。採用了此方法後,本論文不只在 ModelNet40 資料集上達到了最先進的正確率 93.7%,在 ScanObjectNN 資料集上的錯誤率相比於 DGCNN 也減少了 2.96% ~ 7.49%。第二個 方法則是動態 K 值調整 (Dynamic K),我們藉由動態調整 K-近鄰演算 法 (KNN) 的大小來改善在面對低解析度物體時的正確率。有了這個方 法後,我們在面對低解析度物體時,正確率有著 2.4% ~ 434.7% 增長。 最後第三種方法我們利用了神經網路搜索 (NAS) 的技術來找出更適合 點雲分類任務的架構。經由實驗結果證明,神經網路搜索 (NAS) 的方 法確實能帶來更好的性能。透過此方法,我們在 ModelNet40 的正確率 進一步提升到了 93.9%,在 ScanObjectNN 的正確率也與人工設計的架 構表現相當。 In recent years, the investment in automatic driving technology has led to rapid growth of 3D point cloud researches. Due to the irregular and unordered properties of 3D point cloud, it is very difficult to capture the geometric features between the points. In this thesis, we propose three methods to improve the ability of capturing point cloud features to improve the accuracy and stability of the point cloud classification task. In the first approach, we introduce two different attention mechanisms: the Point-wise Attention Module, which determines the correlation between points, and the Channel-wise Attention Module, which allows the model to focus more on important features under limited resources. With these attention mechanisms, we not only achieve the state-of-the-art accuracy of 93.7% on the ModelNet40 [36] dataset, but also reduce the error rate ranging from 2.96% to 7.49% on the ScanObjectNN [32] dataset compared to DGCNN. The second method is Dynamic K. We dynamically adjust the size of the KNN to improve the accuracy for low resolution objects. By using this method, we have seen a 2.4% to 434.7% increase in accuracy when dealing with low resolution objects. Finally, the third method utilizes the Neural Architecture Search technique to find a more suitable architecture for the point cloud classification task. The experimental results prove that the neural architecture search method does bring better performance. The proposed NAS method further improves the accuracy to 93.9% on ModelNet40 [36], while on ScanObjectNN [32], the accuracy was comparable to that of the handcrafted architecture. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78640 |
DOI: | 10.6342/NTU202004324 |
全文授權: | 有償授權 |
電子全文公開日期: | 2023-10-31 |
顯示於系所單位: | 電子工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-0511202017081900.pdf 目前未授權公開取用 | 6.67 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。