Skip to content
←Back to Open Source

OPEN SOURCE DEEP DIVE

Drone MappingPhotogrammetry3D reconstructionPoint clouds

OpenDroneMap: Turn Aerial Imagery into 3D Maps

An open-source drone photogrammetry pipeline that reconstructs photos or video into point clouds, textured 3D models, orthophotos and DSMs/DTMs ready for GIS workflows.

OpenDroneMap/ODM6.6kPythonAGPL-3.09 min read

An auditable spatial-data pipeline, not a one-shot 3D generator

OpenDroneMap, or ODM, is an open-source photogrammetry pipeline for aerial imagery. It takes an overlapping image set and carries it through feature extraction, camera pose estimation, sparse reconstruction, dense depth estimation, point-cloud filtering, meshing, texturing, georeferencing, elevation modeling and orthophoto generation. The resulting assets are designed for GIS, point-cloud viewers, three-dimensional editors and downstream mapping workflows, rather than existing only inside a closed hosted service.

The project presents itself as a command-line toolkit for Windows, macOS and Linux, with Docker images and a desktop distribution maintained separately. That interface matters for robotics teams. The same processing core can run in a terminal, sit behind a NodeODM service, or be divided across machines by ClusterODM. ODM is therefore data-production infrastructure: a repeatable way to turn a flight or inspection session into geometric products.

Its inputs are broader than photographs from a dedicated survey camera. JPEG, TIFF and DNG images are supported, and MP4, MOV, LRV and TS videos can be placed directly in the dataset. Video is converted into still frames before reconstruction. This makes the tool relevant to consumer drones, handheld capture rigs and inspection robots whose original record is a continuous video rather than a carefully triggered photo sequence.

Frame selection is part of the reconstruction problem

The video path does not simply dump every frame into OpenSfM. Each frame is inspected for blur, black frames and excessive similarity to the previously accepted frame. Rejected frames never consume feature-matching or dense-reconstruction capacity. Accepted frames are resized to a configured maximum dimension and written with image metadata. When a matching SRT file is available, parsed position, altitude and camera information can be inserted into EXIF so that the resulting frames participate in georeferencing and initialization.

More frames are not automatically better. Photogrammetry benefits from viewpoint diversity and stable overlap. A long hover with hundreds of nearly identical images adds computation without adding useful geometric constraints. ODM limits extraction to 500 frames per video by default and exposes the output frame dimension as another control. Those settings determine how much visual evidence reaches reconstruction and how quickly the project consumes memory and disk.

The dataset stage then normalizes the raw capture into a reconstruction-ready collection. It discovers images and masks, reads image databases, validates geolocation files, applies camera projection overrides and handles GPS dilution-of-precision values. Ground control points can replace or augment EXIF georeferencing. Optional AI masks can remove sky or background before matching. For multispectral captures, the pipeline can select a primary reconstruction band, align the other bands and convert digital numbers into reflectance. The point is to resolve camera and geospatial assumptions before expensive reconstruction begins.

The main geometry pipeline

OpenSfM forms the sparse reconstruction backbone. It extracts features, matches images, builds tracks, estimates camera motion and triangulates scene structure. ODM then exports camera calibrations and moves the reconstruction into a geographic coordinate reference system when geolocation or ground control is available. Camera intrinsics, GPS observations, image masks and control points all influence this stage. Without a geographic reference, reconstruction can still produce a coherent local model, but its products do not acquire real-world coordinates automatically.

OpenMVS performs dense depth estimation, depth-map fusion and mesh reconstruction. ODM attempts the accelerated path when a supported GPU is available, falls back to CPU processing when the GPU run fails, and retries with tiled fusion when an out-of-memory termination is detected. Those fallbacks are important in production because the largest aerial jobs are often limited by accelerator memory rather than by a lack of available compute.

The dense point cloud is filtered and optionally clipped to a boundary before it becomes a final output. Statistical filtering removes sparse outliers, while a radius-based sampling option can reduce duplicate density. Optional point classification separates ground from non-ground returns for terrain products. Ground rectification can reclassify incorrect ground points and fill gaps. ODM can export point clouds as LAS or LAZ and can also produce CSV, EPT and COPC representations for interchange or streamed access.

Meshing converts the filtered cloud into a triangulated surface. Texturing projects the undistorted camera images back onto that geometry and writes a textured model. ODM can create the usual multi-file OBJ output, a single-texture OBJ, a binary glTF asset, or a simpler two-and-a-half-dimensional surface for orthophoto production. The geo-referenced textured model remains tied to the survey coordinate frame. This is observed geometry reconstructed from camera views, not an unconstrained asset invented by a generative model.

Orthophotos, DSMs and DTMs answer different questions

The orthophoto stage reprojects perspective images onto a common ground plane, removes displacement caused by camera tilt and terrain relief, then blends imagery and resolves seams. ODM can emit a GeoTIFF, a rendered PNG, a KMZ for globe viewers, static tiles and cutline products. Resolution is expressed in centimeters per pixel, but the requested value is capped by the estimated ground sampling distance. Lowering the number in the command cannot recover detail that the flight did not capture.

A digital surface model represents the visible top surface, including buildings, vegetation and equipment. A digital terrain model attempts to represent bare earth after ground classification. They are not interchangeable elevation rasters. ODM interpolates each product from the filtered point cloud and supports gap filling, cropping, Euclidean distance maps, tiles and Cloud Optimized GeoTIFF output. A construction-volume calculation, a terrain-following route and a city-scale surface model should not silently assume the same elevation semantics.

Together, the orthophoto and elevation raster provide the two common views of a surveyed area: imagery for interpretation and height for geometry. The point cloud preserves free-view geometry, while the textured mesh supports surface visualization and scene review. Keeping these products in one coordinate system reduces the drift introduced when separate tools repeatedly reproject and rescale intermediate assets.

Georeferencing and repeatable execution

ODM can initialize a reconstruction from image GPS or from ground control points. GPS-only processing is convenient and often sufficient for relative mapping, while control points constrain absolute position and help manage large-area deformation. The software includes control points in optimization and reports residuals. For high-precision GPS, the assumed accuracy can be overridden; vertical offsets can be applied for a different height datum. These controls are useful, but they do not excuse inconsistent input coordinate systems.

The pipeline can start at a selected stage, stop after a selected stage, or rerun from a selected stage. Existing intermediate products are reused where possible, and changes to relevant options trigger new work. This makes a failed large job debuggable at the point of failure instead of forcing a full restart from raw images. It also lets an operator retain expensive sparse reconstruction while iterating on point-cloud quality or orthophoto settings.

Very large projects can be split into spatial submodels with an overlap buffer. Each block is reconstructed independently, then point clouds, orthophotos and elevation products are merged. Split size and overlap distance are configurable, and a cluster endpoint can distribute submodels across several machines. The engineering requirement is not merely to finish each block; every submodel must share a consistent geospatial frame and expose enough overlap for a clean merge.

Deployment, interfaces and capacity planning

The shortest deployment path is the official Docker image with a project directory mounted into the container. Images live under a named dataset, and the command selects the project path and dataset. Outputs are organized by processing stage: sparse reconstruction, filtered point cloud, mesh, texturing, georeferenced point cloud and orthophoto. A network-facing service can wrap ODM with NodeODM. ClusterODM can distribute block processing when one workstation is no longer sufficient.

GPU support is real but bounded. ODM can accelerate SIFT feature extraction on supported hardware, while dense reconstruction, mesh processing and raster generation use different libraries and different resource profiles. Capacity planning should therefore focus on image count, image resolution, overlap, point-cloud quality and temporary disk, not on GPU memory alone. The dense stage is often the first place where memory pressure becomes visible.

Point-cloud quality increases processing time sharply, and higher reconstruction settings can multiply the work required for matching, depth estimation and fusion. Orthophoto and elevation resolution are constrained by ground sampling distance, while mesh size is capped by vertex and octree controls. A sensible workflow validates coverage, georeferencing and camera settings at medium quality before spending hours on maximum settings. Turning every parameter up is not a substitute for correct flight planning or adequate overlap.

Why it matters for aerial robots and embodied systems

A drone is a mobile sensing platform. ODM is the offline layer that converts its observations into a durable geometric record of the environment. The resulting point cloud or elevation raster can serve as a map base, the orthophoto can support inspection records and target review, and the textured mesh can feed visualization or simulation. ODM does not replace onboard flight control, real-time localization or obstacle avoidance; it produces the higher-fidelity spatial asset after the collection run.

That separation is useful. A robotics stack can keep time-critical control on the vehicle, while a scheduled ODM job turns the returned media into a versioned site model for planning, acceptance checks, change detection or training-data preparation. For teams that revisit the same site, the command-line interface can be queued and automated without uploading raw imagery to a third-party reconstruction service.

The project is licensed under AGPL-3.0 and integrates OpenSfM, OpenMVS, PDAL, GDAL, GRASS GIS, Entwine, PoissonRecon and MVS Texturing. Its value comes from the integration work that connects camera models, dense matching, geospatial coordinates, point-cloud formats and raster products into one operational pipeline. Teams that need real aerial observations converted into controllable spatial infrastructure, while retaining deployment and data ownership, are the natural users.

As an Amazon Associate, we earn from qualifying purchases.