Pre-training Point Cloud Compact Model with Partial-aware Reconstruction

TL;DR

Point-CPR tackles two practical limitations of masked point modeling for 3D point clouds: positional leakage in decoder inputs and large model sizes. It introduces partial-aware reconstruction, which decouples center coordinates from masked patch inputs, and a compact encoder based on local aggregation to reduce parameters and compute. The method is validated on ShapeNet pre-training and downstream tasks including object classification, detection, part segmentation, and completion, achieving strong performance while using around 2% of the parameters of leading MPM models and significantly lower FLOPs. This approach enables robust, efficient pre-training suitable for deployment on resource-constrained devices such as embedded robotics and AR/VR systems, while delivering competitive or superior performance across diverse 3D understanding tasks.

Abstract

The pre-trained point cloud model based on Masked Point Modeling (MPM) has exhibited substantial improvements across various tasks. However, two drawbacks hinder their practical application. Firstly, the positional embedding of masked patches in the decoder results in the leakage of their central coordinates, leading to limited 3D representations. Secondly, the excessive model size of existing MPM methods results in higher demands for devices. To address these, we propose to pre-train Point cloud Compact Model with Partial-aware \textbf{R}econstruction, named Point-CPR. Specifically, in the decoder, we couple the vanilla masked tokens with their positional embeddings as randomly masked queries and introduce a partial-aware prediction module before each decoder layer to predict them from the unmasked partial. It prevents the decoder from creating a shortcut between the central coordinates of masked patches and their reconstructed coordinates, enhancing the robustness of models. We also devise a compact encoder composed of local aggregation and MLPs, reducing the parameters and computational requirements compared to existing Transformer-based encoders. Extensive experiments demonstrate that our model exhibits strong performance across various tasks, especially surpassing the leading MPM-based model PointGPT-B with only 2% of its parameters.

Figures (5)

• Figure 1: Comparison of (a) vanilla MPM reconstruction and (b) our partial-aware reconstruction. Our reconstruction does not require the center coordinates of masked patches as input. The encoding process is omitted in the figure.
• Figure 2: Accuracy-parameters tradeoff on ScanObjectNN. Our Point-CPR performs best. Please refer to Section \ref{['subsec:exp']} for details.
• Figure 3: The pipeline of our Point-CPR. Given a point cloud, we first encode unmasked features by our compact encoder. Then, we concatenate random masked queries with the encoded features and feed them into our partial-aware decoder to decode the masked portion. Finally, we perform dual reconstruction of semantic and positional coordinates of masked patches.
• Figure 4: The structure of our compact encoder layer and our compact encoder consists of $n$ stacked compact encoder layers.
• Figure 5: The structure of our partial-aware prediction module.