GaussianMamba: Geometric and Relational State Space Scanning of 3D Gaussian Primitives
Organizations: Hacettepe University, Ankara, Turkiye
Abstract
3D semantic occupancy prediction requires representing complex spatial structures and modeling contextual relationships across a scene. 3D Gaussian primitives provide a sparse and adaptive representation of scene geometry and semantics, while Mamba offers efficient sequence modeling for capturing dependencies among primitives. However, Gaussian primitives are irregular and unordered, making sequence construction challenging, as conventional spatial ordering may overlook their anisotropic geometry and feature-semantic relationships. In this paper, we present GaussianMamba, a dual-branch Mamba framework that explicitly incorporates Gaussian-specific geometric and feature-semantic relationships into complementary scanning strategies. The first branch refines a locality-preserving traversal using anisotropic Gaussian support derived from primitive scale and orientation. The second branch reorders spatially organized groups according to feature-semantic affinity and adjusts local features using geometry-weighted semantic differences. The resulting sequences are processed by separate Mamba modules, and their features are fused through a support-aware mechanism that incorporates local feature discrepancies arising from the distinct scanning paths. Experiments on nuScenes-SurroundOcc and SSCBench-KITTI-360 demonstrate consistent improvements over GaussianFormer and GaussianFormer-2 in both IoU and mIoU.
Figures & tables
| Method | IoU | mIoU | barrier | bicycle | bus | car | const. veh. | motorcycle | pedestrian | traffic cone | trailer | truck | drive. suf. | other flat | sidewalk | terrain | manmade | vegetation |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BEVFormer [ 16 ] | 30.50 | 16.75 | 14.22 | 6.58 | 23.46 | 28.28 | 8.66 | 10.77 | 6.64 | 4.05 | 11.20 | 17.78 | 37.28 | 18.00 | 22.88 | 22.17 | 13.80 | 22.21 |
| TPVFormer [ 10 ] | 11.51 | 11.66 | 16.14 | 7.17 | 22.63 | 17.13 | 8.83 | 11.39 | 10.46 | 8.23 | 9.43 | 17.02 | 8.07 | 13.64 | 13.85 | 10.34 | 4.90 | 7.37 |
| TPVFormer* [ 10 ] | 30.86 | 17.10 | 15.96 | 5.31 | 23.86 | 27.32 | 9.79 | 8.74 | 7.09 | 5.20 | 10.97 | 19.22 | 38.87 | 21.25 | 24.26 | 23.15 | 11.73 | 20.81 |
| OccFormer [ 30 ] | 31.39 | 19.03 | 18.65 | 10.41 | 23.92 | 30.29 | 10.31 | 14.19 | 13.59 | 10.13 | 12.49 | 20.77 | 38.78 | 19.79 | 24.19 | 22.21 | 13.48 | 21.35 |
| SurroundOcc [ 27 ] | 31.49 | 20.30 | 20.59 | 11.68 | 28.06 | 30.86 | 10.70 | 15.14 | 14.09 | 12.06 | 14.38 | 22.26 | 37.29 | 23.70 | 24.49 | 22.77 | 14.89 | 21.86 |
| GaussianFormer [ 11 ] | 29.83 | 19.10 | 19.52 | 11.26 | 26.11 | 29.78 | 10.47 | 13.83 | 12.58 | 8.67 | 12.74 | 21.57 | 39.63 | 23.28 | 24.46 | 22.99 | 9.59 | 19.12 |
| Method | Input | IoU | mIoU | car | bicycle | motorcycle | truck | other-veh. | person | road | parking | sidewalk | other-grnd | building | fence | vegetation | terrain | pole | traf.-sign | other-struct. | other-object |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MonoScene [ 3 ] | C | 37.87 | 12.31 | 19.34 | 0.43 | 0.58 | 8.02 | 2.03 | 0.86 | 48.35 | 11.38 | 28.13 | 3.22 | 32.89 | 3.53 | 26.15 | 16.75 | 6.92 | 5.67 | 4.20 | 3.09 |
| Voxformer [ 14 ] | C | 38.76 | 11.91 | 17.84 | 1.16 | 0.89 | 4.56 | 2.06 | 1.63 | 47.01 | 9.67 | 27.21 | 2.89 | 31.18 | 4.97 | 28.99 | 14.69 | 6.51 | 6.92 | 3.79 | 2.43 |
| TPVFormer [ 10 ] | C | 40.22 | 13.64 | 21.56 | 1.09 | 1.37 | 8.06 | 2.57 | 2.38 | 52.99 | 11.99 | 31.07 | 3.78 | 34.83 | 4.80 | 30.08 | 17.51 | 7.46 | 5.86 | 5.48 | 2.70 |
| OccFormer [ 30 ] | C | 40.27 | 13.81 | 22.58 | 0.66 | 0.26 | 9.89 | 3.82 | 2.77 | 54.30 | 13.44 | 31.53 | 3.55 | 36.42 | 4.80 | 31.00 | 19.51 | 7.77 | 8.51 | 6.95 | 4.60 |
| GaussianFormer [ 11 ] | C | 35.38 | 12.92 | 18.93 | 1.02 | 4.62 | 18.07 | 7.59 | 3.35 | 45.47 | 10.89 | 25.03 | 5.32 | 28.44 | 5.68 | 29.54 | 8.62 | 2.99 | 2.32 | 9.51 | 5.14 |
| GaussianFormer + Ours † | C | 37.87 | 13.95 | 18.81 | 1.32 | 5.02 | 20.17 | 7.34 | 3.80 | 49.29 | 12.20 | 27.87 | 5.14 | 31.27 | 6.32 | 31.95 | 8.30 | 3.39 | 2.82 | 11.06 | 5.03 |
| Configuration | #B | CGSS | DRSS | GSDF | IoU | mIoU |
|---|---|---|---|---|---|---|
| GaussianFormer [ 11 ] | 0 | 29.83 | 19.10 | |||
| Single Mamba | 1 | 30.05 | 19.18 | |||
| Dual Mamba † | 2 | 30.24 | 19.23 | |||
| + GSDF | 2 | 30.27 | 19.25 | |||
| + CGSS | 2 | 30.88 | 19.48 | |||
| + DRSS | 2 | 30.68 | 19.62 |
| Scanning Strategy | #B | IoU | mIoU | Latency (ms) |
|---|---|---|---|---|
| XYZ | 1 | 29.96 | 19.06 | 6.99 |
| Morton | 1 | 30.00 | 19.13 | 6.10 |
| Hilbert | 1 | 30.05 | 19.18 | 10.68 |
| Hilbert + Hilbert | 2 | 30.24 | 19.23 | 16.16 |
| Hilbert + Reversed Hilbert | 2 | 30.29 | 19.28 | 16.74 |
| Hilbert + XYZ | 2 | 30.48 | 19.42 | 17.40 |
| Geometry-Branch Order | Contact (%) | IoU | mIoU |
|---|---|---|---|
| Hilbert | 12.60 | 30.68 | 19.62 |
| CGSS w/o Gaussian support | 13.40 | 31.21 | 19.91 |
| CGSS | 18.70 | 31.82 | 20.25 |
| Group routing | Boundary correction | IoU | mIoU |
|---|---|---|---|
| 30.98 | 19.55 | ||
| 31.39 | 19.97 | ||
| 31.20 | 19.80 | ||
| 31.82 | 20.25 |
| Fusion Strategy | Params (K) | IoU | mIoU | Latency (ms) |
|---|---|---|---|---|
| Addition | 16.51 | 31.54 | 19.96 | 0.61 |
| Concatenation | 32.80 | 31.82 | 20.25 | 0.82 |
| Cross-attention | 155.32 | 32.01 | 20.29 | 9.13 |
| GSDF † | 33.20 | 32.41 | 20.55 | 2.80 |
| Method | #G | Params (M) | Mem. (GB) | Lat. (ms) | IoU | mIoU |
|---|---|---|---|---|---|---|
| GaussianFormer | 144K | 75.55 | 2.77 | 354.71 | 29.83 | 19.10 |
| GaussianFormer + Ours † | 144K | 76.35 | 2.79 | 466.12 | 32.41 | 20.55 |
| GaussianFormer | 25.6K | 57.09 | 2.08 | 183.51 | 28.72 | 16.00 |
| GaussianFormer + Ours † | 25.6K | 57.89 | 2.11 | 214.12 | 29.60 | 16.97 |
3D Occupancy Prediction
3D occupancy prediction estimates the geometry and semantics of occupied and free space around an agent from camera, LiDAR, or radar data.
See all 125 papers