- Rating
- 2
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
- T-FastGS extends FastGS to reconstruct clean static Gaussian scenes from multi-view images containing transient objects, targeting both reconstruction robustness and training efficiency.
- Its Mask-Guided Error-Aware Density Control (MGDC) separates reliable static errors from transient interference, preserving densification opportunities in transient regions while excluding them from pruning-score computation.
- Asynchronous Alternating Optimization (AAO) updates the mask predictor less frequently and reuses cached masks, reducing overhead while Gaussian optimization continues.
- Experiments on NeRF On-the-go and RobustNeRF report rendering quality comparable to existing transient-robust methods, with substantially fewer Gaussians and approximately 3–3.5× faster training than RobustSplat. Ablations evaluate the contributions of MGDC and AAO.
Critical strengths
- Mask-based training suppresses transient interference, while AAO reduces the cost of maintaining these masks by avoiding redundant updates. Its ablation shows a 38.9% reduction in T-FastGS training time with negligible changes in rendering quality (Table 4).
- MGDC thoughtfully integrates masks into FastGS's density-control mechanism. Its asymmetric design preserves densification opportunities for occluded static structures while excluding transient errors from pruning scores, addressing a limitation of simply masking the reconstruction loss (Sec. 3.3 and Table 3).
- The method achieves a strong quality–efficiency trade-off across two benchmarks. On NeRF On-the-go, it delivers rendering quality comparable to RobustSplat with approximately 3.5× faster training and 0.20M rather than 0.98M Gaussians (Appendix Table 9).
Critical weaknesses
- The MGDC ablations do not isolate the contributions of its densification and pruning strategies or establish the benefit of their asymmetric combination. Separate component ablations, as provided in FastGS, are missing (Tables 3 and 5).
- AAO lacks an ablation of mask-update frequency. The nearly unchanged rendering quality with and without AAO in Table 4 leaves the choice of update interval and the quality–efficiency trade-off across different frequencies insufficiently justified.
- The reported training time may exclude DINOv2 feature precomputation for the input images, which occurs before timing begins in the supplied code. This leaves the end-to-end runtime and fairness of the reported speedup unclear.
Questions for authors
- Could you ablate different mask-update intervals in AAO, reporting rendering quality and training time to justify the chosen frequency and demonstrate the effectiveness of the schedule?
- Could you separately ablate MGDC's densification and pruning strategies, including their combination, to quantify each component's contribution to rendering quality, Gaussian count, and training efficiency?
- Could you report the DINOv2 feature-precomputation time and total runtime including preprocessing, using a consistent timing protocol for all methods?
- I would be willing to raise my score if the authors satisfactorily address these concerns.