基於可變形模板匹配之弱監督三維物體檢測

Abstract

三維物體偵測是三維視覺的一個熱門研究領域，近年來受到廣泛關注。然而，訓練用於三維物體偵測的深度學習模型通常需要大量帶有三維邊界框註釋的數據，這是一項耗時的任務並且存在重大挑戰。為了應對這一挑戰，我們提出了一種通過可變形模板匹配（DTMNet）進行弱監督三維物體偵測的方法，該方法在圖像和二維實例遮罩的弱監督下，通過將可變形形狀模板與輸入的LiDAR點雲進行匹配，生成弱監督的三維虛擬邊界框。生成的三維虛擬邊界框可以用於訓練基於圖像或基於LiDAR的三維物體偵測器。我們的DTMNet顯著降低了註釋成本，提高了三維物體偵測的效率。對KITTI基準數據集的實驗結果在定量和定性上證明了我們提出的模型的有效性和實用性

3D object detection is an active research topic for 3D vision and has been widely studied in recent years. However, training deep learning models for 3D object detection typically requires extensive data with 3D bounding box annotations, which is a time-consuming task and presents a significant challenge. To address this challenge, we propose a weakly supervised 3D object detection method via deformable template matching (DTMNet), which generates weakly supervised 3D pseudo-bounding boxes by matching a deformable shape template with the input LiDAR point clouds under the weak supervision of images and 2D instance masks. The generated 3D pseudo-bounding boxes can be used to train either image-based or LiDAR-based 3D object detectors. Our DTMNet significantly reduces annotation costs and improves the efficiency of 3D object detection. Experimental results on the KITTI benchmark dataset quantitatively and qualitatively demonstrate the effectiveness and practicality of our proposed model.