Skip to content

Latest commit

 

History

History
55 lines (36 loc) · 3.55 KB

troubleshooting.md

File metadata and controls

55 lines (36 loc) · 3.55 KB

Troubleshooting

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.

Hardware requirements

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

GPU / OpenCL selection

  • 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_device to false to allow them.
  • There is no option to pin a specific device index.
  • No OpenCL devices found — cannot compute depthmaps means no usable GPU was detected; the dense stages cannot run.

Dense: running out of memory (RAM or VRAM)

The single biggest driver of memory use and point count is the processed image resolution. In order of impact:

  1. Lower depthmap_max_image_size (default 3200). This is the main lever — fewer points, less RAM and VRAM.
  2. Halve depthmap_fusion_svo_max_voxels (default 80000000) and depthmap_max_cluster_views (default 48) to cap the fusion/cleaning peak memory.
  3. Set depthmap_fusion_svo_voxel_level to half or quarter — coarser voxels mean faster processing and lower memory, at the cost of detail.

Dense: point-cloud density

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

Dense: cropping the output

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.

Dense: skipping LAS/LAZ export

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.

See also