yxjiang0125/mvs

How yxjiang0125/mvs works

MVSAnywhere estimates metric depth from a reference image and posed source views, with demo/inference entry points for scans and user recordings. Dataset adapters supply multi-view imagery and camera geometry; the depth model builds multi-view features and predicts depth, with optional monocular calibration and refinement. Demo options can fuse depth into meshes. The repository also supports training and evaluation, and a separate RegSplatfacto integration uses MVSAnywhere depth for Gaussian-splatting regularization. The README establishes these capabilities; sampled excerpts are partial, so uncertain internal wiring is omitted.

Main entry points: sr_depth_model.py Depth Model · run_demo.py Demo Runner · train.py Training Entry · datasets Dataset Adapters

Components 22

Data and Geometry

Dataset Adapters src/mvsanywhere/datasets
MVS Tuple Dataset src/mvsanywhere/datasets/generic_mvs_dataset.py
Geometry Utilities src/mvsanywhere/utils/geometry_utils.py

Depth Estimation

Depth Model src/mvsanywhere/experiment_modules/sr_depth_model.py
Cost Volume src/mvsanywhere/modules/cost_volume.py
Feature Volumes src/mvsanywhere/modules/feature_volume.py
Network Blocks src/mvsanywhere/modules/networks.py
Metric Calibration src/mvsanywhere/modules/metric_calibration.py

Inference and Fusion

Demo Runner src/mvsanywhere/run_demo.py
Inference API hubconf.py
Depth Fusion src/mvsanywhere/tools/partial_fuser.py
TSDF Volume src/mvsanywhere/tools/tsdf.py
Mesh Rendering src/mvsanywhere/tools/mesh_renderer.py

Training and Evaluation

Depth Losses src/mvsanywhere/losses.py
Training Entry src/mvsanywhere/train.py
Evaluation scripts/evaluation.py

Splat Regularization

Splat Data Manager src/regsplatfacto/regsplatfacto/data/regsplatfacto_datamanager.py
MVS Splat Dataset src/regsplatfacto/regsplatfacto/data/mvsanywhere_dataset.py
RegSplatfacto Model src/regsplatfacto/regsplatfacto/regsplatfacto_model.py
Splat Rendering src/regsplatfacto/regsplatfacto/render_for_meshing.py

Other components

Scan / Recording
Depth / Mesh Output

Connections 24

Scan / Recording → Demo Runner: provides scan
Demo Runner → Dataset Adapters: configures input
Dataset Adapters → MVS Tuple Dataset: adapts tuples
Demo Runner → Depth Model: runs inference
Inference API → Depth Model: invokes model
Depth Model → Cost Volume: builds matching
Depth Model → Feature Volumes: uses features
Depth Model → Network Blocks: uses blocks
Depth Model → Metric Calibration: calibrates depth
Depth Model → Geometry Utilities: uses geometry
Depth Model → Depth / Mesh Output: predicts depth
Demo Runner → Depth Fusion: requests fusion
Depth Fusion → TSDF Volume: integrates depth
TSDF Volume → Mesh Rendering: extracts mesh
Mesh Rendering → Depth / Mesh Output: produces mesh
Training Entry → Dataset Adapters: loads training data
Training Entry → Depth Model: trains model
Training Entry → Depth Losses: computes losses
Evaluation → Depth Model: evaluates predictions
Splat Data Manager → MVS Splat Dataset: creates dataset
MVS Splat Dataset → Inference API: requests depth
Splat Data Manager → RegSplatfacto Model: supplies training data
RegSplatfacto Model → Splat Rendering: renders splats
Splat Rendering → Depth / Mesh Output: produces render

Drawn by GitDiagram on September 29, 2026 from the repository's file tree, README and source files. Read as Markdown ·