基於輕量化基礎變換模型之顏色輔助點雲密度增強方法

Abstract

點雲資料在三維任務中扮演著重要角色，但由於傳感器的硬體限制和環境因素，收集的點雲可能會稀疏或部分缺失，這給後續處理帶來了困難。因此，點雲密度增強方法在點雲資料處理領域中顯得尤為重要。過往研究提出的方法多基於點雲的幾何特徵進行操作和預測。本研究提出了CART (Color-Aided Registration Transformer) 點雲疊合模型，該模型結合了幾何和顏色特徵，能有效預測來源與目標點雲的重疊部分並預測轉換矩陣、對低重疊率的點雲進行精確疊合，從而增強點雲密度、以提升物件辨識的精度。 在真實物體掃描的資料集 Google Scanned Objects 上，CART 模型達到了2.7度的旋轉誤差和0.66公分的平移誤差，與未使用顏色輔助的模型 RegTR 相比，在僅增加不到1%的模型尺寸的情況下，旋轉和平移誤差降低了20%至60%。此外，在大型 3D CAD 模擬資料集 ShapeNet 中，CART 模型達到了3.55度和0.77公分的誤差。特別是在來源和目標點雲存在大角度旋轉（如90度和180度）差異的情況下，CART 模型由於顏色特徵的輔助，對於對稱物體不受影響，極大幅度地降低了誤差，對比 RegTR 模型，誤差降低了80%。

Point cloud data plays a crucial role in 3D tasks, but due to the limitations of sensors and environmental factors, the collected point clouds are usually sparse or partially occluded, making subsequent processing difficult. Therefore, point cloud density enhancement methods are significant for point cloud data processing. Previous studies have mostly operated and predicted the transformation based on geometric features of point clouds. This study proposes a point cloud registration model called CART (Color-Aided Registration Transformer), which combines geometric and color features. It use the overlapping parts between source and target point clouds and then predicts the transformation, effectively registering point clouds with low overlap and enhancing point cloud density. On the scanned object dataset, Google Scanned Objects, the CART model achieved a rotation error of 2.7 degrees and a translation error of 0.66 cm. Compared to the RegTR model, which does not use color assistance, the CART model reduced the rotational and translational errors by 20% to 60%, with an increase in model size of less than 1%. Additionally, on the large-scale 3D CAD simulation dataset, ShapeNet, the model limited the errors up to 3.55 degrees and 0.77 cm. Furthermore, in cases where there is a large rotational difference, such as 90 or 180 degrees, between the source and target point clouds, the CART model, aided by color features, significantly reduces the error for symmetric objects. Compared to RegTR, the error was reduced for up to 80%.