Model Zoo

February 17, 2026 ยท View on GitHub

SceneFlow

ModelOriginal PaperOursConfigurationTrain Input SizeUniform
STTR0.430.40sttr_flyingthings3d.yamlbatchrandom-
PSMNet1.090.95psmnet_sceneflow.yaml256x5120.93
CFNet1.040.96cfnet_sceneflow.yaml256x5120.89
AANet0.870.80aanet_sceneflow.yaml288x5760.84
MobileStereo-2D1.141.10msnet2d_sceneflow.yaml256x5120.99
MobileStereo-3D0.800.78msnet3d_sceneflow.yaml256x5120.75
GwcNet0.760.74gwcnet_sceneflow.yaml256x5120.66
COEX0.680.67coex_sceneflow_amp.yaml288x5760.64
FADNet++0.760.65fadnet_sceneflow.yaml384x768-
CasStereo0.720.67casnet_psm_sceneflow.yaml256x5120.58
IGEV0.470.46igev_sceneflow_amp.yaml256x5120.46
IGEV++0.430.44igevpp_sceneflow_amp.yaml256x7680.38
IGEV_RT0.500.52igev_rt_sceneflow_amp.yaml320x7680.52
StereoBase (Ours)-0.34stereobase_sceneflow.yaml320x7360.34
FoundationStereo0.340.34foundationstereo_sceneflow.yaml320x7360.34
Monster0.370.40monster_sceneflow.yaml320x7360.33

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.

ModelEPEConfigurationTrain Input Size
LightStereo-S (Ours)0.73lightstereo_s_sceneflow.yaml256x512
LightStereo-M (Ours)0.62lightstereo_m_sceneflow.yaml256x512
LightStereo-L (Ours)0.59lightstereo_l_sceneflow.yaml256x512
LightStereo-H (Ours)0.51lightstereo_lx_sceneflow.yaml256x512
StereoBase (Ours)0.34stereobase_sceneflow.yaml320x736

KITTI2015 Leadboard

ModelD1-bgD1-fgD1-all
PSMNet1.864.622.32
PSMNet*1.804.582.26
IGEV1.382.671.59
IGEV*1.442.311.59

The model without * indicates the results of the original paper. The model with * means the implementation of OpenStereo.

ModelD1-bgD1-fgD1-all
LightStereo-S (Ours)2.003.802.30
LightStereo-M (Ours)1.813.222.04
LightStereo-L (Ours)1.782.641.93
LightStereo-H (Ours)1.602.921.82
StereoBase (Ours)1.282.261.44

Access the checkpoint: BaiduDrive or Google Drive or Huggingface