A test splat after training for OVER 5 HOURS on an iPhone17 Pro.
- Captured with my PoC iOS app, which stores Colmap compatible sparse data for each frame.
- For this test splat I did futther feature matching off the device using Colmap.
- Then re-imported the data back into my app for 3500 iterations at 1280px max image size.
- The training took nearly six hours! But completed. (I also pointed a fan at it, since it can get pretty toasty, especially while plugged in to keep charge up.)
- I then brought it into here for a lot of clean up.
All in all a good start.
---
The mobile training tests I'm doing are based on Swift and Metal code I've been coaxing Claude Fable to come up with, after pointing it to the 'PocketGS' paper. (https://arxiv.org/abs/2601.17354). I'm currently attempting to implement it as an iOS Capacitor Plugin, in order to potentially add it as a feature to my capture app.
Still lots of work to do! (If I can afford the tokens!)