- Timeline
- 2026
- Team
- BD3D research group
- My role
- Pipeline, feature engineering, and modeling
- Skills
- Python, Agisoft Metashape, ICP, Ridge regression, LOOCV
The goal
Photogrammetry turns a stack of photos into a 3D model, but a full reconstruction is expensive and you only learn whether it worked at the end. The question: can signals from the earliest stages of the pipeline predict the final reconstruction quality before you commit the compute?
Highlights
- 01
A ground truth for 'good'
I built an Agisoft Metashape pipeline and scored each finished reconstruction against a reference surface using ICP alignment. That gave every run a quality number to predict, and image-overlap graphs from the input photos gave the features to predict it with.
- 02
Early signals beat the obvious baseline
With leakage-aware Ridge models (no information from the finished reconstruction allowed into the features), early structure-from-motion features cut leave-one-out RMSE by 13.7% compared with a reprojection-error baseline.
−13.7%LOOCV RMSE vs. baseline
- 03
An honest negative result
The improvement didn't hold under extrapolation, meaning scenes outside the range the model was trained on. Reporting that clearly matters as much as the headline number: it says the features capture something real, but not yet something general.
Looking forward
The next step is camera-coverage features: describing how thoroughly the cameras actually see each part of the object, which should transfer across scenes better than overlap statistics.