Question about Gaussian Splatting behavior in different scene scales

[1397201241]

Description:

Hi, thank you for your great work.

I’m a bit confused about the behavior of Gaussian Splatting during reconstruction in two different experiments, and would appreciate your insights.

Experiment 1:

• Command:
  frgs reconstruct /xxx/data/xxx/MEDIA_and_1205_ALL_sfm_fastmap_vocab_20251217143531/undistorted -o /xxx/data/xxx/MEDIA_and_1205_ALL_sfm_fastmap_vocab_20251217143531/undistorted/fvdb_r2/output.ply --tx.image-downsample-factor 2 --cfg.max-epochs 50

• Details:

  • num_images: 4271
  • iterations: 4271 * 50 = 213,550
  • image-downsample-factor: 2 (3947 × 2960 → 1973 × 1480)
  • num_gaussians_before: 4,055,534
  • num_gaussians_after: 3,203,845

• Observation:
  The scene is large, starting with a high number of Gaussians, but the count decreased after training.

Experiment 2:

• Command:
  frgs reconstruct /xxx/data/100MEDIA_sfm_20251209082948 -o /xxx/data/100MEDIA_sfm_20251209082948/frgs_r2.ply --tx.image-downsample-factor 2 --cfg.max-epochs 100

• Details:

  • num_images: 858
  • iterations: 858 * 100 = 85,800
  • image-downsample-factor: 2 (3947 × 2960 → 1973 × 1480)
  • num_gaussians_before: 823,733
  • num_gaussians_after: ~11,190,000 (11.19M)

• Observation:
  This scene is a subset/region of the first one. It starts with fewer Gaussians, but after training the number increased dramatically to over 10 million.

My confusion:
In the first (larger) scene, Gaussians decreased, while in the second (smaller subset) scene, Gaussians grew exponentially. I would expect both to behave similarly in terms of Gaussian pruning/growth given the same downsampling factor.

Could you help me understand what might cause this difference? Is it related to scene complexity, density initialization, or the optimization dynamics in different scales?

Thank you in advance for your time and explanation!