Shorter Splatting Eval

shorter-splatting-eval

Speeding Up the Learning of 3D Gaussians with Much Shorter Gaussian Lists — Liu et al. (2026) (arXiv:2603.09277, 2026)

What this evaluates

Evaluates the training efficiency and reconstruction fidelity of a 3D Gaussian Splatting method that uses scale reset and entropy-constrained alpha blending to reduce Gaussian list lengths.

Datasets

• Mip-NeRF 360 — total ?; splits: train (-1), test (-1)
• Deep Blending — total ?; splits: train (-1), test (-1)
• Tanks and Temples — total ?; splits: train (-1), test (-1)

Metrics

• PSNR (primary) — range: [0, ∞)
  • Peak Signal-to-Noise Ratio computed in linear RGB space between rendered and ground-truth images. Higher is better.
• SSIM — range: [0, 1]
  • Structural Similarity Index Measure comparing local patterns of pixel intensities. Higher is better.
• LPIPS — range: [0, 1]
  • Learned Perceptual Image Patch Similarity using deep features. Lower is better.
• Time(s) — range: seconds
  • Total training time in seconds required to reach the specified iteration count.
• N_G — range: count
  • Total number of 3D Gaussians in the final optimized scene representation.

Input / output format

Input: Multi-view images and corresponding camera poses for a static 3D scene.

Output: Rendered images from the optimized 3D Gaussian representation at specified test viewpoints.

Scoring recipe

def evaluate(rendered_imgs, gt_imgs, training_time_s, final_gaussian_count):
    psnr = compute_psnr(rendered_imgs, gt_imgs)
    ssim = compute_ssim(rendered_imgs, gt_imgs)
    lpips = compute_lpips(rendered_imgs, gt_imgs)
    return {
        'PSNR': psnr,
        'SSIM': ssim,
        'LPIPS': lpips,
        'Time(s)': training_time_s,
        'N_G': final_gaussian_count
    }

Common pitfalls

• Comparing training times without controlling for iteration count or target Gaussian count, as speedups are highly dependent on these constraints.
• Misinterpreting LPIPS directionality (lower is better) compared to PSNR/SSIM (higher is better).
• Ignoring that the paper evaluates under two distinct setups (Taming-3DGS target counts vs. Mini-Splatting2 target counts) which yield different performance baselines.

Evidence (verbatim from paper)

Table 1: Quantitative comparison across three datasets. We follow Taming-3DGS’s target Gaussian count setup (3.3M, 2.8M, 1.8M for the three datasets respectively) for fair comparison. | Method | Iters | Mip-NeRF 360 | | | | | Deep Blending | | | | | Tanks and Temples | | | | | | | | N_G | PSNR↑ | SSIM↑ | LPIPS↓ | Time(s)↓ |

Citation

@misc{liu2026shortersplatting,
  title={Speeding Up the Learning of 3D Gaussians with Much Shorter Gaussian Lists},
  author={Liu et al. (2026)},
  year={2026},
  note={arXiv:2603.09277}
}

• arXiv: 2603.09277