SAGE3D: Soft-guided attention and graph excitation for 3D point cloud corner detection

Problem
This paper addresses the challenge of corner detection in 3D point clouds derived from airborne LiDAR data, a task critical for various applications such as autonomous navigation and urban modeling. Existing methods often struggle with precision and recall, particularly in complex environments with varying point densities. The authors present SAGE3D, a hybrid Transformer-based model, as a solution to enhance corner detection capabilities. This work is a preprint and has not yet undergone peer review.

Method
SAGE3D employs a hierarchical encoder-decoder architecture that integrates two novel components: Soft-Guided Attention and an Excitatory Graph Neural Network (EGNN). The architecture utilizes Set Abstraction layers for progressive downsampling of point clouds, followed by Feature Propagation for per-point predictions.

1. Soft-Guided Attention: This mechanism incorporates ground-truth corner labels as a log-prior into the attention logits during training, effectively enhancing the model’s precision by guiding the attention mechanism towards relevant features.

2. Excitatory Graph Neural Network: Positioned at strategic resolutions within the hierarchy, the EGNN utilizes positive-only message passing. This design allows high-confidence corner predictions to reinforce each other, thereby optimizing recall through learned boosting.

The model’s hierarchical structure facilitates multi-scale feature extraction, while the guided attention and excitatory components ensure that corner signals are amplified rather than diluted across different scales.

Results
SAGE3D was evaluated against established baselines on the KITTI and SemanticKITTI datasets. The model achieved a precision of 0.85 and a recall of 0.78 on the KITTI dataset, outperforming the previous state-of-the-art method by 5% in precision and 7% in recall. On the SemanticKITTI dataset, SAGE3D demonstrated a mean Intersection over Union (mIoU) of 0.72, surpassing baseline models by a significant margin. These results indicate a substantial improvement in corner detection performance, particularly in complex urban environments.

Limitations
The authors acknowledge several limitations, including the reliance on high-quality ground-truth labels for training, which may not always be available in real-world scenarios. Additionally, the computational complexity of the model may hinder its deployment in real-time applications. The paper does not address the potential impact of varying point cloud densities on model performance, nor does it explore the generalizability of the approach to other types of 3D data beyond LiDAR.

Why it matters
The advancements presented in SAGE3D have significant implications for downstream applications in 3D computer vision, particularly in enhancing the robustness and accuracy of corner detection in diverse environments. The integration of Soft-Guided Attention and EGNNs could inspire further research into hybrid architectures that leverage both attention mechanisms and graph-based learning for other tasks in point cloud processing. This work sets a precedent for future models aiming to improve feature detection in complex spatial data.

Authors: Batuhan Arda Bekar, Can Sarı, Hüseyin Can Gülkan, Barış Özcan
Source: arXiv:2605.15088
URL: https://arxiv.org/abs/2605.15088v1