- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies hyperspectral novel-view synthesis when each training view observes only a sparse subset of spectral bands. Instead of storing an independent C-dimensional spectrum for every Gaussian, SB-HyperGS represents spectral appearance with scene-level shared spectral bases and per-Gaussian mixing coefficients, performs splatting in the low-dimensional material space, and decodes the result to arbitrary wavelengths. It further adds spectral smoothness and mixture sparsity regularization, as well as a capacity-bounded adaptive spatial refinement. Experiments on three synthetic and three real scenes compare the method with Gaussian and NeRF baselines, and with a Direct-HSI control under matched parameter budgets and matched Gaussian counts.
Critical strengths
-
The controlled comparison with Direct-HSI isolates the spectral representation and gives clear evidence that shared spectral bases are valuable under sparse-band supervision. On real scenes with the same number of Gaussians, SB-HyperGS uses roughly half the parameters and still improves PSNR by 1.55–1.80 dB; on synthetic scenes it reduces parameters by about 74% at a cost of only 0.18–0.44 dB. This is the main evidence for the proposed representation and the main reason I see value in the paper.
-
Rendering in the 48-dimensional material space before spectral decoding yields a compact and fast model, with about 2M parameters and more than 430 FPS for full-spectrum rendering. This efficiency is practically important for scaling hyperspectral 3D reconstruction.
Critical weaknesses
-
The real-scene results do not fully support the central capacity-reallocation explanation. On synthetic scenes, the matched-budget advantage shrinks as the budget grows, which is consistent with the claim. On real scenes, however, PSNR decreases as the number of Gaussians increases for both representations, and the full model in Table 2 is below the fixed-topology configuration with 8192 Gaussians in Table 3. This suggests that on real data the gain may mainly come from the low-rank spectral prior of the shared bases rather than from reallocated spatial capacity, and the role of adaptive spatial refinement there is unclear.
-
The incremental contribution over existing compact spectral representations is not established. HyperGS renders hyperspectral Gaussians through a learned low-dimensional latent space, and UnMix-NeRF represents hyperspectral radiance with a global endmember dictionary and spatial abundances, which is close to the proposed shared-bases-plus-local-coefficients design. Both are cited on page 3 but neither is compared in Tables 1–4. In addition, TensoRF-HSI and Mip-NeRF 360-HSI obtain higher PSNR than SB-HyperGS in Table 1, which the text does not discuss.
Questions for authors
-
Could you provide a real-scene ablation of adaptive spatial refinement, and a control that separates capacity reallocation from the low-rank spectral prior? For example, Direct-HSI with a low-rank or spectral-smoothness constraint, or per-Gaussian low-dimensional coefficients without shared bases.
-
Could you report HyperGS and/or an adapted UnMix-NeRF under the same sparse-band observation protocol?
-
Do the Table 2 full model and the fixed-topology runs in Table 3 share the same training settings apart from refinement? What is the Gaussian capacity cap of the full model, and what is the final number of Gaussians for each row of Table 5?