Instance-Aware Kernel Prediction for Efficient Point Cloud Learning

Abstract

As a pivotal modality for 3-D perception, point clouds are essential for various applications ranging from autonomous driving to immersive 3-D reconstruction. However, robust and efficient point cloud representation learning is often hindered by the limitations of static, instance-agnostic neighborhood construction. Existing architectures predominantly rely on Euclidean proximity for local feature aggregation, which inevitably leads to feature contamination at semantic boundaries and prohibitive computational overhead when modeling fine-grained geometric details. To address these challenges, we propose the instance-aware kernel prediction network (IKPNet), a novel framework that integrates semantic consistency into hierarchical representation learning. Specifically, we develop a lightweight score head to generate a semantic affinity matrix that is progressively refined across network layers. This matrix enables dynamic, instance-aware neighborhood construction, ensuring that feature aggregation is governed by both geometric proximity and semantic coherence. By sharing and iteratively updating this affinity matrix, IKPNet facilitates the capture of nonlocal semantic correlations while significantly reducing redundant multi-layer perceptron (MLP) computations. Extensive experiments on multiple benchmarks demonstrate that IKPNet achieves state-of-the-art performance, including 94.1% classification accuracy on ModelNet40 and 74.9% mean intersection over union (mIoU) on S3DIS Area 5, while maintaining computational efficiency. Our results indicate that bridging the gap between low-level geometry and high-level semantic consistency is crucial for precise and scalable 3-D scene understanding.