Pitch · one sentence + numbers
You say “I built the multi-camera capture system and the datasets, then compared photogrammetry vs NeRF vs 3DGS on the same inputs — including glass and white plastic where classical pipelines fail.”
Dataset 15 objects · deliberate hard materials included.
Dig deeper → Studio protocol · Status matrix
System · acquisition stack
You say “Not only training configs — twelve industrial cameras, BLE turntable, C++/C# capture software, session layout for offline pipelines.”
Physical 12-cam rig.
CamMatrixCapture UI · ~8k LOC.
WinUI → CaptureCore → Camera / BLE / Session.
Dig deeper → Architecture · Capture loop · Hardware bandwidth
Results · same capture, two methods
You say “Identical multi-view input. Splatfacto averages ~33 dB, Nerfacto ~22 dB on studio. Watch the metal pot — specular moves with the camera.”
Metal pot
Metal pot
Method matrix from the thesis.
Full gallery → Results gallery
Hard materials · why the dataset matters
You say “We put failure cases in the set on purpose. Glass and white plastic break feature matching. Neural methods still synthesize views. Cross-pol can rescue classical pipelines on glossy surfaces.”
Glass · Splatfacto (photo fails geometry).
Without cross-pol — specular mess.
With cross-pol — matching can recover.
Example photogrammetry mesh plate.
Dig deeper → Evaluation · hard materials
Outdoor · scale + takeaway
You say “Gränsö Castle: ~5.2k images, 77% aligned at 0.61 px. Photogrammetry still wins for full geometry; neural wins for bounded high-quality viz. Hybrid workflow.”
RealityCapture aerial / site context.
Splatfacto outdoor still.
RC flythrough.
Close the loop
Dataset · software · PDF — hand them links and stop re-explaining.
Dig deeper → Outdoor chapter · What to claim / not claim