基於擴散模型之條件式樹形點雲生成

Abstract

在現實環境中，樹木因其自然生長的特性，具有高度不規則性與遮蔽性，使得以傳統 3D 掃描技術完整擷取其三維結構變得極具挑戰。特別是樹冠部分由眾多形狀與方向各異的葉片構成，且樹木通常具一定高度，若僅以 3D 掃描器進行掃描，往往僅能獲取局部且未被遮擋的點雲資訊，導致樹木的 3D 重建結果片段且不完整。本論文針對上述問題，提出一套基於擴散模型（Diffusion Model）的條件式 3D 樹形點雲生成方法。我們以不完整的樹形點雲作為條件輸入，並透過擴散模型學習特定樹種中樹幹、樹枝與樹冠的幾何分佈與結構，並且我們在訓練階段，提出了一個基於樹木骨架的加權採樣機制，在生成結果上顯示，此採樣方法能很好的保持樹幹樹枝的連續性，我們也將傳統只有坐標值的點雲生成擴展到具顏色的點雲生成，生成結果保留原始輸入中的可見資訊，生成符合該樹種特性且具有顔色、結構合理性的完整樹形點雲。隨後，我們對生成點雲進行上採樣，進一步提升點雲的密度與視覺真實感，最終透過 EWA Surface Splatting 進行算圖，將處理過後的樹形點雲進行可視化呈現。實驗結果顯示，本方法能按照所輸入的不完整樹形點雲，生成合理、完整且多樣的樹木結構。In real-world environments, trees exhibit high degrees of irregularity and occlusion due to their natural growth characteristics, making it extremely challenging to fully capture their 3D structures using traditional 3D scanning techniques. This is especially true for the canopy, which consists of numerous leaves with varying shapes and orientations. Moreover, trees are typically tall, and when scanned using only 3D scanners, the result often includes only partial and unoccluded point cloud data. Consequently, the reconstructed 3D models of trees tend to be fragmented and incomplete.To address this issue, this thesis proposes a conditional 3D tree-shaped point cloud generation method based on diffusion models. We use incomplete tree point clouds as conditional inputs, and train a diffusion model to learn the geometric distribution and structure of trunks, branches, and canopies specific to different tree species. During training, we introduce a skeleton-based weighted sampling mechanism, which demonstrates strong continuity in trunk and branch generation. Furthermore, we extend aditional point cloud generation (which only includes coordinates) to support colorized point cloud generation. The generated outputs retain visible information from the input while producing complete tree-shaped point clouds that align with the characteristics of the given tree species in terms of color and structural plausibility.After generation, we perform upsampling on the point clouds to increase their density and visual realism. Finally, we apply EWA Surface Splatting for rendering, enabling highquality visualization of the processed tree point clouds. Experimental results show that our method can generate reasonable, complete, and diverse tree structures conditioned on the input incomplete point clouds.