MambaSplat-4D

Weights for MambaSplat-4D (ACCV 2026), a model that classifies sequences of 3D Gaussian Splats (bird, drone, airplane, helicopter) with predictions that are rotation-invariant by construction, even when every frame is rotated independently. Data: AeroSplat-4D.

Paper · Project page · Code

z-trained checkpoints, best seed per run; the paper reports the seed mean (held-out z/z accuracy, %).

file	paper table	seed	acc.	paper
mambasplat_X_z.pt	main Tab. 1, 91 assets	42, best of 3	81.7	80.7 ± 1.2
mambasplat_C_z.pt	main Tab. 1, 91 assets	42, best of 3	80.4	76.5 ± 5.8
open/mambasplat_X_z.pt	supp. Tab. 1, 52 public assets	1, best of 5	80.4	75.8 ± 4.1
open/mambasplat_C_z.pt	supp. Tab. 1, 52 public assets	2, best of 5	67.3	58.8 ± 10.2

X: all 14 Gaussian attributes; C: centres only (dataset.feature_mode=C). Evaluate on the public test split with open/ only; the 91-asset models trained on 5 of its 13 test assets.

Citation

@inproceedings{verdiesen2026mambasplat4d,
  title     = {{MambaSplat-4D}: Rotation-Invariant {4D} Classification of {Gaussian} Splatting Sequences},
  author    = {Verdiesen, Alessandro and Hofstee, H. Peter and Al-Ars, Zaid},
  booktitle = {Asian Conference on Computer Vision (ACCV)},
  year      = {2026}
}