姚遥-Matrix3D.pdf

1、摄影测量模型Matrix3D:Large Photogrammetry Model All-in-One姚遥南京学 NJU-3DVIntroduction|Photogrammetry Pipeline(a)Multi-view images(b)Camera parameters(c)Dense point cloudSfMMVSNVSMeshingDesigning a Unified Model for 3D ReconstructionApplicable to multiple sub-tasks(e.g.,SfM,MVS,NVS)Introduction|3D Content Cr

2、eationMulti-viewReconstructionColmapMVSNetNeRF3DGSDust3rObservationsDesigning a Unified Model for 3D Content CreationApplicable to multiple sub-tasks(e.g.,SfM,MVS,NVS)Introduction|3D Content CreationMulti-viewSingle-viewReconstructionColmapMVSNetNeRF3DGSDust3rGenerationClayDirect3DCraftsManTripoSGTr

3、ellisObservationsDesigning a Unified Model for 3D Content CreationApplicable to multiple sub-tasks(e.g.,SfM,MVS,NVS)For both reconstruction and generationMatrix3D|Large Photogrammetry Model All-in-OneGenerative Architectures:Multi-modal multi-view Diffusion TransformerPhotogrammetry tasks as modalit

4、y conversionMask learning for multi-task trainingUnified RepresentationsImages:RGB imagesCameras:Ray mapsGeometry:depth mapsMatrix3D|Pose EstimationInput SettingImagesNoisy camerasMatrix3D|Pose EstimationInput SettingImagesNoisy camerasMatrix3D|Multi-view Depth EstimationInput SettingImagesCamerasNo

5、isy depth mapsMatrix3D|Novel View SynthesisInput SettingImagesCamerasNovel cams and noisy imagesMatrix3D|Novel View SynthesisInput SettingImagesCamerasNovel cams and noisy imagesMatrix3D|Unposed Few-shot ReconstructionOne model,the whole pipeline Matrix3D|Unposed Few-shot ReconstructionMatrix3D|Unposed Few-shot ReconstructionMatrix3D|Unposed Few-shot ReconstructionMatrix3D|Large Photogrammetry Model All-in-One