The tables in this article show the memory consumption peaks and spent time for the common processing operations for different types of surveys: nadir aerial survey and close-range survey in order to give the general understanding of the time and RAM required for regular workflow procedure.
The processing has been performed in Metashape Professional on compute nodes with the following hardware configuration (using a single node for each operation):
| CPU | 32 vCPU (2.7 GHz Intel Xeon E5 2686 v4) |
|---|---|
| GPU | 2 x NVIDIA Tesla M60 |
| RAM | 240 GB |
Aerial data processing benchmark
The dataset used to benchmark aerial data processing consists of 920 images, 40 MPix resolution each. The data has been acquired using Wingtra platform (WingtraOne VTOL mapping drone) with Sony RX1RII camera, coordinates measured with a precise PPK method. No GCPs data was used for processing.
For evaluation purposes the dataset can be downloaded from the Wingtra website ("High-resolution quarry mapping" dataset):
Wingtra: https://wingtra.com/mapping-drone-wingtraone/aerial-map-types/#3d-reconstruction
The alignment has been performed using Medium and High accuracy in turns with Generic+Reference preselection and 40,000 / 4,000 key/tie point limits for both cases. Adaptive camera model fitting - disabled in each processing scenario.
All further processing operations have been performed based on the alignment corresponding to High accuracy matching.
Image matching, Depth maps calculation, and depth maps based mesh and DEM generation operations used GPU acceleration.
Table 1. Aerial data processing benchmark in Metashape Professional 1.7.2:
| Task (Source)/ Parameters | Time | Memory Peak |
|---|---|---|
| Match Photos | ||
| Medium accuracy | 23 m 02 s | 758 Mb |
| High accuracy | 22 m 46 s | 603 Mb |
Align cameras | ||
| (Medium accuracy) | 7 m 27 s | 416 Mb |
| (High accuracy) | 7 m 08 s | 427 Mb |
Depth maps | ||
| High/Mild | 3 h 08 m | 20.41 Gb |
| Medium/Mild | 47 m | 8.51 Gb |
| Low/Mild | 19 m 23 s | 5.99 Gb |
| Point Cloud (Dense cloud) | ||
| High/Mild | 2 h 36 m | 15.03 Gb |
| Medium/Mild | 37 m | 13.28 Gb |
| Low/Mild | 12 m 33 s | 7.51 Gb |
| Mesh (from Depth Maps) | ||
| High/Mild | 10 h 12 m | 14.80 Gb |
| Medium/Mild | 2 h 15 m | 13.04 Gb |
| Low/Mild | 43 m | 10.80 Gb |
| DEM (from Point Cloud (Dense Cloud)) | ||
| High | 16 m 28 s | 350 Mb |
| Medium | 5 m | 335 Mb |
| Low | 2 m | 331 Mb |
| DEM (from Depth Maps) | ||
| High | 3 h 37 m | 12.18 Gb |
| Medium | 1 h 06 m | 8.13 Gb |
| Low | 25 m 26 s | 5.59 Gb |
| Orthomosaic (DEM) | ||
| DEM (High) | 3 h 05 m | 4.89 Gb |
| DEM (Medium) | 1 h 30 m | 8.96 Gb |
| DEM (Low) | 1 h 16 m | 11.94 Gb |
| Tiled Model (from Depth maps, texture on GPU) | ||
| High | 43 h | 15.05 Gb |
| Medium | 9 h 19 m | 10.49 Gb |
| Low | 4 h 38 m | 6.81 Gb |
Memory consumption for the orthorectification process (the first part of the Build Orthomosaic stage) is proportional to the number of CPU threads.
Close-range data processing benchmark
The dataset used for close-range processing performance benchmarking has been captured with DJI Phantom 4 RTK drone and consists of 124 overhead flight images (nadir and oblique) and 648 images taken by the automated mission plan following. In total 772 images of 18 MPix were used.
The overview of the reconstructed mesh model with the camera locations and working volume is shown on the following screenshot:
The alignment has been performed using Medium and High accuracy in turns with Generic preselection only and 40,000 / 4,000 key/tie point limits for both cases. Adaptive camera model fitting and Exclude stationary points options - disabled in each processing scenario.
All further processing operations have been performed based on the alignment corresponding to High accuracy matching.
Image matching, Depth maps calculation, and depth maps based mesh generation operations used GPU acceleration.
Table 2. Close-range data processing benchmark in Metashape Professional 1.7.2:
| Task (Source) / Parameters | Time | Memory Peak |
|---|---|---|
| Match Photos | ||
| Medium accuracy | 30 m 43 s | 466 Mb |
| High accuracy | 36 m 21 s | 522 Mb |
| Align cameras | ||
| (Medium accuracy) | 8 m 24 s | 785 Mb |
| (High accuracy) | 10 m 27 s | 880 Mb |
| Depth maps | ||
| High/Mild | 43 m | 8.02 Gb |
| Medium/Mild | 14 m | 3.72 Gb |
| Low/Mild | 7 m | 2.38 Gb |
| Point Cloud (Dense Cloud) | ||
| High/Mild | 1 h 24 m | 25.88 Gb |
| Medium/Mild | 24 m 13 s | 14.94 Gb |
| Low/Mild | 10 m 31 s | 5.34 Gb |
| Mesh (from Depth Maps) | ||
| High/Mild | 2 h 59 m | 11.22 Gb |
| Medium/Mild | 1 h 8 m | 8.89 Gb |
| Low/Mild | 26 m 33 s | 9.16 Gb |
| Texture UV (Generic mapping) | ||
| High (32.2 M faces) | 51 m | 7.28 Gb |
| Medium (8.9 M faces) | 33 m | 6.12 Gb |
| Low (2.4 M faces) | 22 m | 6.78 Gb |
| Texture Blending (on GPU only, 16K) | ||
| Medium (8.9 M faces ) | 22 m 50 s | 10.94 Gb 6 Gb VRAM |
| Low (2.4 M faces ) | 14 m 44 s | 10.05 Gb 6 Gb VRAM |
| Texture Blending (on CPU only, 16K) | ||
| High (32.2 M faces) | 48 m | 33.86 Gb |
| Medium (8.9 M faces) | 24 m | 32.95 Gb |
| Low (2.4 M faces) | 17 m | 32.86 Gb |
| Tiled Model (from Depth Maps, texture on CPU) | ||
| High/Mild | 7 h 43 m | 27.86 Gb |
| Medium/Mild | 4 h 8 m | 28.14 Gb |
| Low/Mild | 3 h 22 m | 26.47 Gb |
Memory consumption for the texture blending operation (the second part of the Build Texture stage) is proportional to the number of CPU threads if performed on the CPU. There's no such limitation when texture blending uses GPU acceleration.