- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This work addresses the central challenge of single-global-representation alignment in cross-modal self-supervised 3D point cloud understanding. The authors argue that existing methods overlook the hierarchical semantic structure and spectral properties of point cloud representations, leading to suboptimal transfer performance. To address this, they propose MS-POINT, a hierarchical-spectral structured cross-modal pretraining framework. The method extracts multi-level intermediate features from the 3D backbone (DGCNN by default), performs graph-filter-inspired low/high frequency decomposition on selected stages, and integrates these features via an adaptive fusion module. On the image side, they construct hierarchical representations with coarse/fine decomposition to provide structurally matched supervision. Extensive experiments demonstrate that MS-POINT outperforms previous cross-modal baselines such as CrossPoint and MM-Point across linear classification, few-shot learning, and segmentation tasks.
Critical strengths
The paper explores a potentially interesting direction of combining hierarchical feature learning and spectral analysis for cross-modal 3D pretraining. The experimental evaluation covers multiple standard benchmarks and downstream tasks, providing a basic assessment of the proposed method. The authors also include a series of ablation studies on the components of their framework, which offers some insight into the contribution of each module.
Critical weaknesses
This study has shortcomings in terms of technical rigor and experimental validity: the benchmark comparisons are outdated and incomplete; unsupervised state-of-the-art classical monomodal methods (such as Inter-MAE) have been overlooked; the ablation experiments failed to control for interference from parameters and computational costs, and thus cannot prove that the performance improvement stems from the spectral decomposition itself rather than from additional model capacity; and no comparisons were made with other spectral methods. (The paper is quite extensive; if I have overlooked anything during my review, I suggest the auth
Questions for authors
-
Could you add comparisons with Inter-MAE(TMM'23) and TPM(TCSVT'25) on ScanObjectNN and ModelNet40? If you can demonstrate that MS-POINT significantly outperforms these latest baselines with a reasonable computational cost, I would be willing to reconsider my score.
-
The current ablation studies do not isolate the effect of the spectral decomposition from the increase in parameters. Could you perform a controlled ablation where you compare the proposed graph-filter decomposition with a simple MLP or linear projection that has exactly the same number of parameters and computational cost? This would help determine whether the performance improvement comes from the spectral decomposition itself or just the additional model capacity.
-
The paper only evaluates on object-level tasks and small-scale indoor segmentation. Could you provide results on a large-scale outdoor dataset such as SemanticKITTI for semantic segmentation? This would demonstrate that your method generalizes to more complex real-world scenarios.
-
The authors claim that {A} acts as a low-pass filter and I-{A} acts as a high-pass filter, but they provide no theoretical analysis or references to support this. Could you provide a detailed explanation of why this formulation corresponds to spectral filtering, including a formal connection to the graph Fourier transform? Without this theoretical justification, the spectral decomposition component appears arbitrary and unmotivated.
-
The proposed method increases training time and parameters compared to MM-Point for a 1.8% accuracy improvement on ScanObjectNN. Could you justify this trade-off? Are there any optimizations you can make to reduce the computational cost without sacrificing performance?