Abstract
We present a Non-parametric Network for 3D point cloud analysis, Point-NN, which consists of purely non-learnable components: farthest point sampling (FPS), k-nearest neighbors (k-NN), and pooling operations, with trigonometric functions. Surprisingly, it performs well on various 3D tasks, requiring no parameters or training, especially for few-shot classification. Starting from this basic non-parametric model, we propose two extensions. First, Point-NN can serve as a base architectural framework to construct Parametric Networks by simply inserting linear layers on top. Given the superior non-parametric foundation, the derived Point-PN exhibits a high performance-efficiency trade-off with only a few learnable parameters. Second, Point-NN can be regarded as a plug-and-play module to enhance the already trained 3D models without re-training. Point-NN captures the complementary geometric knowledge and boosts existing methods for different 3D benchmarks by interpolating the predictions during inference. We conduct extensive experiments to revisit the non-parametric network components and demonstrate their significance for 3D point cloud understanding.
Introduction
In the rapidly evolving field of deep learning, the prevailing paradigm has been "bigger is better," characterized by a relentless drive toward scaling up parameters and designing increasingly complex operators. In 3D point cloud analysis, since the seminal PointNet++, the focus has shifted heavily toward "learnable" complexity—adding adaptive convolutions, graph-based message passing, and millions of parameters to handle unordered and irregular data.
However, this trend often obscures the fundamental power of basic geometric operations. This paper asks a provocative question: How much of a 3D model's success is truly derived from its learned weights, versus the underlying geometric framework? To answer this, we introduce Point-NN, a framework that achieves impressive results on 3D tasks with zero parameters and zero training.
The Point-NN Framework
Point-NN is built upon purely non-learnable components, demonstrating that geometry alone is a powerful prior for 3D understanding. The architecture consists of two primary stages: a Non-Parametric Encoder and a Point-Memory Bank.
Non-Parametric Encoder
The encoder transforms raw 3D coordinates into high-dimensional feature vectors using a multi-stage hierarchy. It utilizes Farthest Point Sampling (FPS) for downsampling and k-Nearest Neighbors (k-NN) for local grouping. A key innovation is the use of Trigonometric Positional Encoding (PosE). By applying sine and cosine functions to relative 3D coordinates, the model captures local spatial patterns through simple dot products, effectively replacing the need for learned weights in feature extraction. Features are then expanded and aggregated through standard Max and Average pooling.
Point-Memory Bank
To perform classification without a learnable head, Point-NN employs a Point-Memory Bank. During a "pre-inference" phase, the features of the training set are cached. At inference time, the test sample's features are compared against this memory using cosine similarity. The labels are then determined by a similarity-weighted vote from the cached training samples.
Extensions: Point-PN and Plug-and-Play
Beyond its utility as a standalone zero-parameter model, Point-NN serves as a versatile foundation for two major extensions.
Point-PN: A High-Efficiency Parametric Network
By treating Point-NN as a base architectural framework, we can construct Point-PN (Parametric Network). This is achieved by simply inserting a few linear layers into the existing non-parametric structure. Because the underlying geometric foundation is so strong, Point-PN achieves state-of-the-art performance with a fraction of the parameters required by traditional models.
Plug-and-Play Enhancement
Point-NN can also be utilized as a "plug-and-play" module to boost the performance of already trained 3D models. By interpolating the predictions of a trained network with those of Point-NN during inference, we can capture complementary geometric knowledge. This enhancement requires no additional training and consistently improves accuracy across various benchmarks.
Experimental Evaluation
We evaluated Point-NN and its extensions across standard benchmarks, including ModelNet40, ScanObjectNN, ShapeNetPart, and ScanNetV2.
Key Results
- Zero-Shot Performance: On ModelNet40, Point-NN achieved 81.8% accuracy with zero parameters, outperforming several early fully-trained deep learning models.
- Few-Shot Dominance: Point-NN excels in low-data regimes where trained models often overfit. In 10-way 20-shot tasks, it reached 84.9% accuracy, significantly outperforming PointCNN (50.0%).
- Efficiency of Point-PN: Point-PN reached 93.8% accuracy on ModelNet40 with only 0.8M parameters, demonstrating a superior performance-efficiency trade-off compared to models like PointMLP (12.6M parameters).
- Plug-and-Play Boost: Fusing Point-NN with a pre-trained PointMLP improved its accuracy on the real-world ScanObjectNN dataset by +2.0% without re-training.
Conclusion
Point-NN challenges the necessity of massive parameter counts in 3D point cloud analysis. By demonstrating that a non-parametric model can achieve competitive results, we highlight the importance of geometric priors. Whether used as a standalone zero-parameter baseline, a foundation for efficient parametric models, or a plug-and-play enhancement, Point-NN provides a new perspective on building efficient and robust 3D vision systems. Future work will explore scaling these non-parametric approaches to even more complex, large-scale indoor and outdoor scenes.
Summary updated