Description – brief description of the issue to resolve or decision to make
For our first field test with the Phoenix Pro on 3/7 03/07/25 Fire Academy Visit , we are aiming to port at least 1 algorithm to the drone and test its performance in smoke.
This decision will help focus our development efforts so that we can start testing one subsystem early.
Options – what alternatives were available/considered?
Depth
Mast3r
MoGE
Odometry
ROVTIO
MSO
Resolution – description of the changes or decision made. If not resolved, put a note like "Pending" here.
We decided to go with MoGE.
Justification – supporting information for why this solution or decision was deemed best
- Given the dependency of odometry algorithms on camera-IMU calibrations and camera-imu timesync which are facing implementation roadblocks as of the current time, we decided to go with a depth estimation algorithm to test first.
- Deciding depth algo: currently, unable to achieve feature matching and metric depth with Mast3r, not at same level of maturity as MoGE. However, point-based reconstruction is still worth some investigation wrt l3d class project
Abhishek's l3d project on point-based reconstruction methods (mast3r adjacent)
Swastik's l3d project on human detection on point clouds
We have yet to conclusively decide on metric depth performance but we can benchmark this on Mar 7th (no smoke environment).