Geometrically-Grounded Bounding Box Refinement for 3D Object Detection : 3차원 객체 검출을 위한 기하학 기반 바운딩 박스 정제

다국어 초록 (Multilingual Abstract)

Three-dimensional object detection from point clouds has become a crucial capability for embodied artificial intelligence, supporting perception and in- teraction in robotics, autonomous navigation, and mixed reality. Although recent detectors achieve strong overall performance, they still fail to localize objects that exhibit high structural complexity. Objects such as furniture, ap- pliances, and composite assemblies contain multiple interleaved components and fine geometric details. These heterogeneous structures lead to inconsistent feature aggregation within the same object, which causes fragmented or over- sized bounding boxes. A major contributing factor is the inability of volumet- ric IoU-based supervision to provide stable and continuous gradients—par- ticularly when predicted and ground-truth boxes barely overlap-resulting in unreliable boundary learning in sparse indoor point clouds. To address this limitation, we propose PointRefine3D, a geometry-aware refinement framework that improves the localization precision of existing 3D detectors without modifying their backbones. The framework consists of two complementary components. First, the Point IoU Loss optimizes the align- ment between predicted boxes and observed object points rather than vol- umetric overlap. This point-based formulation produces smooth gradients near object boundaries and enforces surface-level coverage, even for sparsely sampled or partially missing shapes. Because initial box predictions are of- ten unstable in sparse point clouds, an additional refinement mechanism be- comes necessary. Accordingly, we introduce an IoU Regression Head that predicts the localization quality of each detected box and guides selective re- finement. During inference, boxes with low predicted IoU values are refined by an Object-Aware Refinement module that recomputes the box center and size from the spatial distribution of surrounding points while discarding pla- nar outliers. This selective strategy improves boundary accuracy with minimal computational overhead. We also note that our baseline detector (UniDet3D) adopts axis-aligned ground-truth annotations, meaning that all ground-truth yaw angles are θ = 0, which we explicitly account for in our formulation. Experiments on four indoor benchmarks—ScanNetV2, S3DIS, MultiScan, and 3RScan—demonstrate that PointRefine3D consistently outperforms UniDet3D. The method achieves an average gain of +1.8 mAP@0.5 and up to +4.1 on the challenging 3RScan dataset. These results verify that combining point-level geometric supervision with selective refinement yields robust and precise 3D detection, especially for complex, multi-part objects with interleaved geome-

목차 (Table of Contents)

• Abstract 1
• I. Introduction 3
• II. Related Work 7
• 2.1 Point cloud-based 3D object detection 7
• 2.2 Intersection over Union loss in detection 8
• 2.3 Bounding box post-processing and refinement 9
• III. Preliminaries 10
• 3.1 Task Definition 10
• 3.2 Baseline Framework 11
• IV. Methodology 11
• 4.1 Overview 11
• 4.2 Point IoU Loss 13
• 4.3 IoU Regression for Quality Estimation 16
• 4.4 Object-Aware Refinement 17
• 4.4.1 Point Sampling: 17
• 4.4.2 Surface Point Selection: 17
• 4.4.3 Tight Bound Estimation: 17
• 4.4.4 IoU-Adaptive Correction 18
• 4.5 Training and Implementation Details 19
• 4.5.1 Network Configuration 20
• 4.5.2 Hyperparameter Settings 20
• V. Experiments 21
• 5.1 Experimental Setup 21
• 5.2 Comparison with State-of-the-Art 22
• 5.3 Per-Class Performance Analysis 25
• 5.4 Ablation Study 27
• 5.5 Qualitative Results 28
• VI. Limitations 30
• VII. Conclusion 31
• References 32
• 국문 초록 38