[Paper Review] Geometric Back-projection Network for Point Cloud Classification

Shi Qiu, Saeed Anwar+1 more|arXiv (Cornell University)|Nov 28, 2019

3D Shape Modeling and Analysis81 references17 citations

TL;DR

This paper proposes a Geometric Back-projection Network (GBN) for point cloud classification that enhances feature learning by explicitly encoding geometric descriptors in low-level 3D space and using an attentional error-correcting feedback mechanism to refine high-level features. The method achieves state-of-the-art performance on synthetic and real-world benchmarks with improved accuracy and efficiency.

ABSTRACT

As the basic task of point cloud analysis, classification is fundamental but always challenging. To address some unsolved problems of existing methods, we propose a network that captures geometric features of point clouds for better representations. To achieve this, on the one hand, we enrich the geometric information of points in low-level 3D space explicitly. On the other hand, we apply CNN-based structures in high-level feature spaces to learn local geometric context implicitly. Specifically, we leverage an idea of error-correcting feedback structure to capture the local features of point clouds comprehensively. Furthermore, an attention module based on channel affinity assists the feature map to avoid possible redundancy by emphasizing its distinct channels. The performance on both synthetic and real-world point clouds datasets demonstrate the superiority and applicability of our network. Comparing with other state-of-the-art methods, our approach balances accuracy and efficiency.

Motivation & Objective

• Address the challenge of limited geometric feature utilization in existing 3D CNNs for point cloud classification.
• Improve feature representation by explicitly enriching low-level geometric information using geometric point descriptors.
• Enable high-level feature learning to capture local geometric context through an error-correcting feedback mechanism.
• Reduce feature redundancy and improve representation compactness using a channel-wise affinity attention module.
• Demonstrate generalization across multiple 3D vision tasks, including classification, part segmentation, and 3D object detection.

Proposed method

• Apply geometric point descriptors (e.g., normal vectors, curvature) as explicit geometric priors in the input 3D space to enrich low-level point features.
• Introduce an attentional back-projection module that mimics error-correcting feedback: the network compares input and reconstructed features to refine high-level representations.
• Use a skip-connection-based forward-pass feedback structure to avoid training instability, inspired by 2D super-resolution and pose estimation methods.
• Combine max-pooling and shared fully connected layers per local neighborhood to extract both prominent and fine-grained local geometric features.
• Integrate a Channel-wise Affinity Attention (CAA) module that reweights feature channels based on inter-channel correlation to suppress redundancy and emphasize distinct features.
• Construct a fully convolutional backbone with multi-scale feature aggregation for segmentation and detection tasks, using the same GBN architecture as the backbone.

Experimental results

Research questions

• RQ1 Can explicit low-level geometric encoding improve 3D point cloud classification beyond standard coordinate-based features?
• RQ2 How can an error-correcting feedback mechanism enhance high-level feature learning in 3D CNNs without introducing training instability?
• RQ3 To what extent does a channel-wise attention mechanism reduce redundancy and improve feature representation in point cloud networks?
• RQ4 Does the proposed architecture generalize effectively across multiple 3D vision tasks beyond classification?
• RQ5 Can the method achieve a better balance between accuracy and computational efficiency compared to state-of-the-art models?

Key findings

• The proposed GBN achieves a 59.3% mAP@0.25 on the SUN RGB-D V2 dataset for 3D object detection, outperforming PointNet, PointNet++, DGCNN, and RS-CNN.
• On the ShapeNet Part Segmentation dataset, the method achieves an instance mIoU of 85.1%, surpassing most baselines and ranking second only to PointASNL.
• The ablation study confirms that both the geometric descriptor and the attentional back-projection module contribute significantly to performance gains.
• The Channel-wise Affinity Attention (CAA) module reduces feature redundancy and improves representation compactness, as shown by visualization and quantitative analysis.
• The network demonstrates strong generalization, achieving state-of-the-art or competitive results across classification, part segmentation, and 3D detection tasks.
• The method maintains high efficiency, balancing accuracy and computational cost better than many existing state-of-the-art approaches.