- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
CoHSPAD reconstructs a dynamic human and static scene as 3D Gaussians from a raw monocular 1-bit SPAD stream. A differentiable Bayer/Poisson/Bernoulli imaging model supervises Gaussian rendering directly with binary photon observations. Instead of independent per-timestamp poses, SMPL motion is represented by sparse B-spline controls, a shared pose residual MLP, and post-LBS Gaussian deformation, allowing observations across time to constrain one continuous trajectory. MTF-Net fuses ordered long- and short-term photon contexts to produce target-aligned human supervision without fixed-window averaging, while SPAD-aware densification separately handles noisy human and static-scene gradients. The authors introduce CoH-Synthesis and CoH-Real and compare against adapted HUGS and MAD-Avatar, with Restormer used for pseudo-RGB preprocessing. CoHSPAD reports 30.646 average human PSNR, improves held-out PSNR from 26.542 to 28.795, and reduces average SMPL MPJPE from 30.75 to 27.38.
Critical strengths
-
Underexplored sensing gap and physically grounded raw-domain supervision. RGB video suffers motion blur, while individual SPAD frames are sparse; direct Bayer/Poisson/Bernoulli supervision is physically aligned with the sensor.
-
Continuous SMPL trajectory stabilizes sparse timestamp supervision. Shared B-spline controls, pose residuals, and post-LBS deformation let photon observations across time jointly constrain motion.
-
A complete system evaluated beyond synthetic fitting. The paper combines human-scene Gaussians, MTF-Net, SPAD-aware densification, real captures, held-out timestamps, pose refinement, and robustness analysis.
Critical weaknesses
-
The experimental validation is too self-contained to establish the claimed robustness and superiority. The method is evaluated only on the authors' newly introduced CoH-Synthesis and CoH-Real datasets, with five synthetic and five real sequences in total. The synthetic observations are generated using the same Poisson/Bernoulli image-formation model that the optimizer assumes, which introduces a potential model-match bias. The real captures have no independent clean RGB ground truth; the real-data silhouette evaluation references Cutie masks that are also used for supervision, so the authors themselves note that it measures consistency with training masks rather than independent reconstruction accuracy. In addition, the quantitative comparison is limited mainly to adapted HUGS and MAD-Avatar variants, without direct photon-, spike-, event-, or continuous-motion reconstruction baselines or evaluation on independent public benchmarks. Consequently, the reported experiments do not sufficiently demonstrate generalization beyond the authors' own data protocol.
-
Novel-view and geometric reconstruction quality are not evaluated, despite the paper's 3D reconstruction and view-synthesis motivation. The introduction frames monocular dynamic human reconstruction as supporting novel-view synthesis and human animation, but the quantitative image metrics are computed at the fixed input viewpoint. The paper reports PSNR, SSIM, and LPIPS against temporally aligned references, plus SMPL pose errors, but it does not evaluate held-out novel camera views, cross-view consistency, depth accuracy, or independent geometry metrics such as point-to-surface or Chamfer errors. Since the input is monocular and the method relies on SMPL, masks, and Gaussian regularization, high fixed-view rendering quality does not by itself establish a correct 3D human-scene decomposition or a view-generalizable representation. Novel-view and geometric evaluations are therefore essential but missing.
-
The method depends heavily on external priors, while the marginal gains of some proposed components are limited. The claim of reconstruction directly from raw SPAD streams should be interpreted with caution: the pipeline requires GVHMR SMPL initialization, synthetic or Cutie human masks, known or externally estimated fixed camera parameters, human-centered crops, and sensor calibration. The initialization-robustness study perturbs joint rotations only up to five degrees, while the optimized trajectory still has substantial errors, including an average MPJPE of 27.38 and much larger errors on difficult scenes. Moreover, the most distinctive temporal-fusion component improves held-out PSNR by only 0.123 dB over continuous SMPL alone, and B-spline interpolation yields only small gains over linear or SLERP alternatives. Without comparisons to simpler trajectory regularization, alternative continuous-motion parameterizations, and other multi-scale temporal-fusion designs, the advantages of the proposed components over less specialized alternatives are not adequately isolated.
Questions for authors
-
How would CoHSPAD perform on independent public SPAD datasets, or with synthetic observations generated by a simulator whose imaging assumptions differ from the paper’s Poisson/Bernoulli model?
-
Can the authors provide held-out novel-view rendering, cross-view consistency, depth, Chamfer, or point-to-surface evaluations to substantiate the 3D reconstruction claim beyond fixed-view rendering?
-
How sensitive is the method to larger failures in GVHMR initialization, Cutie masks, camera intrinsics, fixed camera pose, and SPAD calibration?
-
Since MTF-Net adds only 0.123 dB over continuous SMPL alone in the core ablation, how does it compare with fixed-window accumulation, temporal smoothing, trajectory regularization, and the multi-scale design of USP-Gaussian under matched settings?
-
For CoH-Real, can the authors evaluate against independent masks, multi-view references, depth sensors, or manual annotations rather than Cutie masks already used for supervision?