Starting with ODM version ``0.6.0`` you can split up very large datasets into manageable chunks (called submodels), running the pipeline on each chunk, and then producing merged DEMs, orthophotos and point clouds. The process is referred to as "split-merge".
Why might you use the split-merge pipeline? If you have a very large number of images in your dataset, split-merge will help make the processing more manageable on a large machine (it will require less memory). If you have many machines all connected to the same network you can also process the submodels in parallel, thus allowing for horizontal scaling and processing thousands of images more quickly.
Split-merge works in WebODM out of the box as long as the processing nodes support split-merge, by enabling the ``--split`` option when creating a new task.
Image calibration is recommended (but not required) for large datasets because error propagation due to image distortion could cause a bowl effect on the models. Calibration instructions can be found at `Calibrate Images <tutorials.html#calibrating-the-camera>`_.
:alt:image of lens distortion effect on bowling of data
:align:center
Bowling effect on point cloud over 13,000+ image dataset collected by World Bank Tanzania over the flood prone Msimbasi Basin, Dar es Salaam, Tanzania.
Splitting a dataset into more manageable submodels and sequentially processing all submodels on the same machine is easy! Just use ``--split`` and ``--split-overlap`` to decide the the average number of images per submodels and the overlap (in meters) between submodels respectively
The grouping can be provided by adding a file named image_groups.txt in the main dataset folder. The file should have one line per image. Each line should have two words: first the name of the image and second the name of the group it belongs to. For example::
ODM can also automatically distribute the processing of each submodel to multiple machines via `NodeODM <https://github.com/OpenDroneMap/NodeODM>`_ nodes, orchestrated via `ClusterODM <https://github.com/OpenDroneMap/ClusterODM>`_.
When connected via telnet, it is possible to interrogate what is happening on the cluster. For example, we can use the command HELP to find out available commands
ClusterODM also includes the option to autoscale on multiple platforms, including, to date, Amazon and Digital Ocean. This allows users to reduce costs associated with always-on instances as well as being able to scale processing based on demand.
* Setup a S3-compatible bucket for storing results.
* Create a configuration file for `DigitalOcean <https://github.com/OpenDroneMap/ClusterODM/blob/master/docs/digitalocean.md>`_ or `Amazon Web Services <https://github.com/OpenDroneMap/ClusterODM/blob/master/docs/aws.md>`_.
You should always have at least one static NodeODM node attached to ClusterODM, even if you plan to use the autoscaler for all processing. If you setup auto scaling, you can't have zero nodes and rely 100% on the autoscaler. You need to attach a NodeODM node to act as the "reference node" otherwise ClusterODM will not know how to handle certain requests (for the forwarding the UI, for validating options prior to spinning up an instance, etc.). For this purpose, you should add a "dummy" NodeODM node and lock it
GCPs are fully supported, however, there needs to be at least 3 GCP points on each submodel for the georeferencing to take place. If a submodel has fewer than 3 GCPs, a combination of the remaining GCPs + EXIF data will be used instead (which is going to be less accurate). We recommend using the ``image_groups.txt`` file to accurately control the submodel split when using GCPs.
Huge props to Pau and the folks at Mapillary for their amazing contributions to OpenDroneMap through their OpenSfM code, which is a key component of the split-merge pipeline. We look forward to further pushing the limits of OpenDroneMap and seeing how big a dataset we can process.
