I’m trying to figure out the most cost effective way to use WebODM to build a small number (3-5) of models for personal use. I’m hoping someone here can help me get a handle on this.
I’m thinking about Lightning, WebODM on an M1 MBA, buying hardware including an NVIDIA card, and spinning up VPS for ClusterODM as needed, but I’m open to other options.
Bits I think are relevant:
I’m a software engineer by trade with deep knowledge of Linux, cloud, and VPS.
This is fun so I consider my time to be free.
The models will be built from 300-600 images covering an outdoor area of about 8,000 sq ft.
A representative area has some trees and one building. The primary complexity is rolling hills.
The work is inconsistent; I work on it for a few weeks and then shelve it for a few months.
I imagine I’m going to have to build this few times to get it right and value faster iteration times.
I have experience with photogrammetry and SLAM but none with WebODM and haven’t done anything relevant in over a decade.
I assume I want WebODM but I’m open to the idea of using ODM via the command line.
Simply in terms of $, what hardware (purchased or “rented”) would you use to run ODM to build these models?
This doesn’t directly answer your question, but out of interest I’m running a task now on my M2 MacBook Air (807 images)using WebODM that I’ve previously run on my production WebODM server for comparison. Once it’s done I’ll share the results here as it may at least give you some indication of what would be possible…
Originally my Docker Desktop settings only had 7.8GB RAM allocated (this MBA has 16GB RAM) and the job failed after a while due to insufficient RAM. I then increased the allocation in Docker Desktop to 15GB and re-ran the job (it already had all 8 CPUs allocated). It then ran successfully* in 03:47:48
On my production server which has 64GB RAM, the same task with the same settings took 03:44:44 (so basically the same - not what I expected). The server has 2 vCPUs (Intel Xeon Gold 6426Y; 8 cores per vCPU) on Nutanix AHV virtual server.
*When I say “successfully” - it ran and completed. For some weird reason though it has cut the resulting orthophoto cleanly in half (I have no idea why - it has reconstructed less than half the points):
