| Model | Original Paper | Ours | Configuration | Train Input Size | Uniform |
|---|---|---|---|---|---|
| STTR | 0.43 | 0.40 | sttr_flyingthings3d.yaml | batchrandom | - |
| PSMNet | 1.09 | 0.95 | psmnet_sceneflow.yaml | 256x512 | 0.93 |
| CFNet | 1.04 | 0.96 | cfnet_sceneflow.yaml | 256x512 | 0.89 |
| AANet | 0.87 | 0.80 | aanet_sceneflow.yaml | 288x576 | 0.84 |
| MobileStereo-2D | 1.14 | 1.10 | msnet2d_sceneflow.yaml | 256x512 | 0.99 |
| MobileStereo-3D | 0.80 | 0.78 | msnet3d_sceneflow.yaml | 256x512 | 0.75 |
| GwcNet | 0.76 | 0.74 | gwcnet_sceneflow.yaml | 256x512 | 0.66 |
| COEX | 0.68 | 0.67 | coex_sceneflow_amp.yaml | 288x576 | 0.64 |
| FADNet++ | 0.76 | 0.65 | fadnet_sceneflow.yaml | 384x768 | - |
| CasStereo | 0.72 | 0.67 | casnet_psm_sceneflow.yaml | 256x512 | 0.58 |
| IGEV | 0.47 | 0.46 | igev_sceneflow_amp.yaml | 256x512 | 0.46 |
| IGEV++ | 0.43 | 0.44 | igevpp_sceneflow_amp.yaml | 256x768 | 0.38 |
| IGEV_RT | 0.50 | 0.52 | igev_rt_sceneflow_amp.yaml | 320x768 | 0.52 |
| StereoBase (Ours) | - | 0.34 | stereobase_sceneflow.yaml | 320x736 | 0.34 |
| FoundationStereo | 0.34 | 0.34 | foundationstereo_sceneflow.yaml | 320x736 | 0.34 |
| Monster | 0.37 | 0.40 | monster_sceneflow.yaml | 320x736 | 0.33 |
Underline refers to evaluation in the non-occluded regions only. Uniform: consistent settings, including identical GPU hardware, execution environments, crop sizes, and the same optimizer and learning rate schedule.
| Model | EPE |
Configuration | Train Input Size |
|---|---|---|---|
| LightStereo-S (Ours) | 0.73 | lightstereo_s_sceneflow.yaml | 256x512 |
| LightStereo-M (Ours) | 0.62 | lightstereo_m_sceneflow.yaml | 256x512 |
| LightStereo-L (Ours) | 0.59 | lightstereo_l_sceneflow.yaml | 256x512 |
| LightStereo-H (Ours) | 0.51 | lightstereo_lx_sceneflow.yaml | 256x512 |
| StereoBase (Ours) | 0.34 | stereobase_sceneflow.yaml | 320x736 |
| Model | D1-bg |
D1-fg |
D1-all |
|---|---|---|---|
| PSMNet | 1.86 | 4.62 | 2.32 |
| PSMNet* | 1.80 | 4.58 | 2.26 |
| IGEV | 1.38 | 2.67 | 1.59 |
| IGEV* | 1.44 | 2.31 | 1.59 |
The model without * indicates the results of the original paper. The model with * means the implementation of OpenStereo.
| Model | D1-bg |
D1-fg |
D1-all |
|---|---|---|---|
| LightStereo-S (Ours) | 2.00 | 3.80 | 2.30 |
| LightStereo-M (Ours) | 1.81 | 3.22 | 2.04 |
| LightStereo-L (Ours) | 1.78 | 2.64 | 1.93 |
| LightStereo-H (Ours) | 1.60 | 2.92 | 1.82 |
| StereoBase (Ours) | 1.28 | 2.26 | 1.44 |