Point Neighborhood Embeddings
Pedro Hermosilla
TL;DR
This paper extensively studies Point Neighborhood Embeddings (PNE) in 3D point cloud neural networks, revealing that simple linear embeddings outperform complex MLP-based ones and guiding improved architecture design.
Contribution
It provides the first controlled experimental analysis of PNE mechanisms, offering recommendations that enhance neural network performance on point cloud tasks.
Findings
MLP-based embeddings with ReLU perform worse than simple linear embeddings.
Simple convolutional architectures with linear embeddings achieve state-of-the-art results.
Recommendations improve the performance of complex convolution operations in point cloud networks.
Abstract
Point convolution operations rely on different embedding mechanisms to encode the neighborhood information of each point in order to detect patterns in 3D space. However, as convolutions are usually evaluated as a whole, not much work has been done to investigate which is the ideal mechanism to encode such neighborhood information. In this paper, we provide the first extensive study that analyzes such Point Neighborhood Embeddings (PNE) alone in a controlled experimental setup. From our experiments, we derive a set of recommendations for PNE that can help to improve future designs of neural network architectures for point clouds. Our most surprising finding shows that the most commonly used embedding based on a Multi-layer Perceptron (MLP) with ReLU activation functions provides the lowest performance among all embeddings, even being surpassed on some tasks by a simple linear…
Peer Reviews
Decision·Submitted to ICLR 2024
1. The review of previous 3D point cloud embeddings is comprehensive and logical. 2. The experiments and analysis on two tasks (classification and segmentation) are careful and in-depth.
The improvement is not obvious compared to existing methods in Table 2 compared to 3. Does this mean the effect of PNE is less important when training data is less? I recommend to experiment on larger-scale 3D datasets for classification, such as Objverse, or other 3D tasks, such as indoor/outdoor 3D object detection. The two experiments on the paper is not sufficient enough to demonstrate the conclusion.
1. The paper is well-written and provides a good analysis of point cloud embeddings. 2. This can help build new algorithms/architectures to improve embeddings. 3. The study provides interesting results. 4. The architecture includes simple modification and not expensive operations like transformers.
1. This paper can be seen as a good experimental study paper which is not up to the level of ICLR. This is like a review paper although in a different direction where it provides extensive study on multiple points. 2. The paper is very weak in novelty and makes some claims without evidence or explanations except results. 3. The work mentions a lot of comparisons between activation functions, MLP vs. KP, etc. However, the whole paper lacks in explaining “why” something is better or worse. For e
1. Their findings and recommendations can benefit future arch design of point clouds 2. They did comprehensive experiments to explore the different design choices and validate their recommendations. 3. The writing is good with detailed introduction and analysis.
-
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
Topics3D Shape Modeling and Analysis · Surface Roughness and Optical Measurements · Optical measurement and interference techniques
MethodsConvolution