Reducing the Gap Between Spatial and Space-Filling Curve Neighbors in 3D Point Cloud Segmentation

The central observation that SFC serialization breaks spatial locality is well illustrated (Fig. 1 and Fig. 2) and correctly anchored to the PTv3 design choice of replacing KNN with serialized neighbors. The DAS Loss formulation in Eq. 1–Eq. 5 is straightforward to implement: for each scene, a threshold $T = D_{mean} + D_{std}$ selects SFC-adjacent pairs, and binary cosine-similarity terms pull same-class features together while pushing different-class features apart. The three-dataset evaluation and the breakdown into $\mathcal{L}_{same}$ and $\mathcal{L}_{diff}$ (Table 5) strengthen the empirical argument.

First, the loss is applied only to immediately consecutive points on the SFC ($p_i$ and $p_{i+1}$), even though PTv3 defines neighbors as the $K$ points before and after a query point. This means DAS Loss ignores the vast majority of neighbor pairs actually used in attention, leaving the core discrepancy unaddressed for $K>1$. Second, the thresholding rule and the cosine-similarity objective are heuristic: the paper offers no analysis for why $D_{mean} + D_{std}$ is optimal, and the lack of a temperature or margin in the contrastive terms makes the objective brittle if feature magnitudes vary. Notably, on DALES the full DAS Loss underperforms the $\mathcal{L}_{diff}$-only variant (77.53 vs. 77.77), yet the authors dismiss this without a clear explanation.

Additionally, the paper claims generality but tests exclusively on Point Transformer V3; no experiments are run on other serialization-based architectures such as OctFormer or FlatFormer, so it is unclear whether the loss is broadly transferable or merely patches PTv3-specific feature distributions. The smaller improvement on ScanNet V2 is attributed to noise, but no diagnostic experiments—such as controlled noise injection or feature-space visualization—are provided to validate that claim.

The quantitative gains are statistically supported by three-run averages with standard deviations, which is more than many architectural papers provide. However, Table 4 compares the authors' own three-run PTv3 baseline against single-run literature values, and the S3DIS comparison omits several recent strong baselines that could narrow the claimed margin. The qualitative visualization in Fig. 4 is anecdotal: yellow boxes highlight improvements but do not quantify error modes or show failure cases. The authors fairly note the smaller ScanNet V2 improvement and attribute it to noise, though they provide no diagnostic experiments to confirm this hypothesis.

The experimental protocol itself is relatively detailed—batch sizes, epochs, learning rates, and augmentation policies are listed—but the paper does not provide a code repository or exact PTv3 configuration files. Because the method is implemented as a loss added to an existing complex pipeline, the absence of source code and random seeds would block exact reproduction. Reporting standard deviations over three runs is helpful, yet without the training scripts and exact dependency versions, independent verification is difficult.