metashape-reconstruction

Metashape Dense Reconstruction via MCP

Overview

Build dense products from aligned cameras in Metashape using MCP tools. This covers everything AFTER alignment: depth maps, dense point cloud, mesh, texture, DEM, and orthomosaic. All MCP tool calls block until complete — no polling, no timeouts.

When to Use

• Building depth maps, point cloud, mesh, or texture in Metashape
• Generating DEM or orthomosaic survey products
• Any processing step after photo alignment is complete

Golden Rules

1. CPU OFF for all dense operations. set_gpu_config(cpu_enable=False) before depth maps, point cloud, meshing, texturing, DEM, orthomosaic. CPU slows GPU operations. CPU is ONLY for alignment (match_photos, align_cameras).
2. NEVER write pipeline scripts. Call each MCP tool individually, check the result, adapt. This is an AGENT workflow.
3. No timeouts. Dense operations can take hours or days. MCP calls block until done.
4. No polling. Don't call get_processing_status in a loop.
5. Save after every step. The MCP tools auto-save, but verify with save_project() after major operations.
6. Check prerequisites. Each step requires the previous step's output. Verify before proceeding.

Standard Dense Pipeline

Step 0: Verify alignment is complete

get_alignment_stats()
# Check: alignment_rate should be >95%
# Check: tie_point_count_valid should be reasonable

Step 1: GPU config — CPU OFF

set_gpu_config(cpu_enable=False)

Step 2: Build depth maps

build_depth_maps(
    downscale=2,          # 1=Ultra, 2=High, 4=Medium, 8=Low, 16=Lowest
    filter_mode="mild",   # mild for complex terrain, moderate default, aggressive for clean scenes
    reuse_depth=True      # Reuse existing depth maps for aligned cameras
)

Quality guide:

• downscale=1 (Ultra): Final production, small projects. Very slow.
• downscale=2 (High): Standard production quality. Good balance.
• downscale=4 (Medium): Quick results, large projects, testing.

Step 3: Build point cloud (optional — needed for classification)

build_point_cloud(
    point_colors=True,
    point_confidence=True
)

Skip if going directly to mesh from depth maps.

Step 4: (Optional) Classify ground points

Only needed for terrain/DEM workflows:

classify_ground_points(
    max_angle=15.0,
    max_distance=1.0,
    cell_size=50.0
)

Step 5: Build mesh

build_model(
    surface_type="arbitrary",    # "arbitrary" for 3D, "height_field" for terrain/DEM
    source_data="depth_maps",    # "depth_maps", "point_cloud", "depth_maps_and_laser_scans"
    interpolation="enabled",     # "disabled", "enabled", "extrapolated"
    vertex_colors=True,
    vertex_confidence=True,
    volumetric_masks=False,
    keep_depth=True
)

• face_count_custom=0 is hardcoded (unlimited faces)
• trimming_radius=0 is hardcoded (no trimming)
• For terrain-only mesh from classified point cloud: classes=[2] (ground only)

Step 6: Build UV

build_uv(
    mapping_mode="generic",    # "generic", "adaptive_orthophoto", "orthophoto", "camera"
    texture_size=8192          # Power of 2: 4096, 8192, 16384
)

Step 7: Build texture

build_texture(
    blending_mode="natural",   # "natural" (default, best quality), "mosaic", "average", "max", "min"
    texture_size=8192,
    ghosting_filter=True
)

Step 8: (Optional) Survey products

DEM:

build_dem(
    source_data="point_cloud",  # or "mesh"
    interpolation="enabled",
    classes=[2]                  # Ground only, if classified
)

Orthomosaic:

build_orthomosaic(
    surface_data="dem",         # or "mesh", "none"
    blending_mode="mosaic",
    ghosting_filter=True
)

Surface Type Selection

Use Case	surface_type	source_data
3D object/scene	arbitrary	depth_maps
Terrain/landscape	height_field	depth_maps
With laser scans	arbitrary	depth_maps_and_laser_scans
From dense cloud	arbitrary	point_cloud
Quick preview	arbitrary	tie_points

Depth Map Filter Modes

Mode	When to Use
mild	Complex terrain, vegetation, rock faces — preserves detail
moderate	Default, good for most scenes
aggressive	Clean/flat surfaces, buildings — removes more noise

Texture Blending Modes

Mode	When to Use
natural	Default. Best color continuity across seams
mosaic	Sharpest detail per-patch but visible seams
average	Smooth blending, can be slightly blurry

Quality vs Speed Tradeoffs

For testing/preview: downscale=4 depth maps, texture_size=4096 For production: downscale=2 depth maps, texture_size=8192 For ultra quality: downscale=1 depth maps, texture_size=16384

Sky/Tunnel Artifact Prevention

Road corridor captures commonly produce tunnel/dome mesh artifacts where the sky gets closed. This is because depth map interpolation ignores masks.

Quick fix: Build mesh from point cloud instead of depth maps:

build_point_cloud(point_colors=True, point_confidence=True)
build_model(source_data="point_cloud", surface_type="arbitrary", classes=[0, 1, 2, 6])
clean_model(criterion="component_size", level=75)

For the full decision tree and all 5 strategies, see the sky-artifact-prevention skill.

Texture Pipeline

For UV mapping, texture atlas settings, blending modes, and texture artifact diagnosis, see the texturing-pipeline skill.

Quick reference:

build_uv(mapping_mode="generic", texture_size=8192)
build_texture(blending_mode="mosaic", texture_size=8192, ghosting_filter=True)

Common Issues

• Out of GPU memory on depth maps: Increase downscale (e.g., 2→4). Ultra (1) requires significant VRAM.
• Holes in mesh: Try interpolation="extrapolated" or build from point cloud instead of depth maps.
• Sky/tunnel artifacts: Build from point cloud, not depth maps. See sky-artifact-prevention skill.
• Blurry texture: Increase texture_size, or try blending_mode="mosaic". See texturing-pipeline skill.
• Mesh too large for export: Use decimate_model(face_count=target) after building.