Quick fixes for the most common issues, focused on the GPU dense stages where hardware limits bite first. All settings below go in your dataset's config.yaml; see the configuration reference for full descriptions.
The dense pipeline (depthmaps, fusion, DSM/ortho) needs an OpenCL-capable GPU. The sparse SfM stages run on CPU, except for the matching, which is GPU OpenCL by default, and can be deactivated by setting matcher_type to FLANN,
| RAM | GPU (OpenCL) | |
|---|---|---|
| Minimum | 16 GB | 8 GB VRAM |
| Preferred | 32 GB | 16 GB VRAM |
- Depthmap estimation (
compute_depthmaps) uses all detected non-Intel OpenCL GPUs in parallel — add a second card and it is used automatically. - Fusion (
fuse_depthmaps) runs on the first GPU (device 0) only. - Intel GPUs are skipped by default (they are slow/buggy for this workload). Set
opencl_ignore_intel_gpu_devicetofalseto allow them. - There is no option to pin a specific device index.
No OpenCL devices found — cannot compute depthmapsmeans no usable GPU was detected; the dense stages cannot run.
The single biggest driver of memory use and point count is the processed image resolution. In order of impact:
- Lower
depthmap_max_image_size(default3200). This is the main lever — fewer points, less RAM and VRAM. - Halve
depthmap_fusion_svo_max_voxels(default80000000) anddepthmap_max_cluster_views(default48) to cap the fusion/cleaning peak memory. - Set
depthmap_fusion_svo_voxel_leveltohalforquarter— coarser voxels mean faster processing and lower memory, at the cost of detail.
depthmap_fusion_svo_decimate_flat trims points on flat surfaces while preserving sharp edges:
| Value | Effect |
|---|---|
1.0 |
No decimation — densest cloud |
2.0 |
Balanced (default) |
4.0 |
More adaptive decimation — sparser flats, detail kept at edges |
By default the dense outputs (point cloud + DSM/ortho) are cropped to the PCA-trimmed convex hull of the SfM ground points, trimming the sparse fringe beyond the surveyed area.
- Crop more aggressively: raise
dense_crop_percentileto5.0or10.0. - Disable cropping entirely: set
dense_crop_to_sfm_hulltofalse.
LAS and LAZ export are on by default (alongside fused.ply). To skip them, set both dense_pointcloud_export_las and dense_pointcloud_export_laz to false.
- Configuration reference — every parameter and its default.
- Dense reconstruction & 2D maps — the dense pipeline and its outputs.
- Workflow presets — sensible starting configs per capture type.