Researchers at Cornell Tech and the Royal Danish Academy won Best Paper at ACADIA'26 for Pentimento, a tool that documents on-site construction changes with gaussian splatting.
The paper, Building Drift, first appeared on arXiv in June and was presented at the conference in Detroit, held October 22 to 24. Ritik Batra, a doctoral candidate in Thijs Roumen's lab at Cornell Tech, is lead author, with co-authors from Cornell and the Royal Danish Academy's Centre for Information Technology and Architecture (CITA).
The authors define building drift as the collective deviation of a structure from its digital model as builders adapt to reclaimed materials. Their case study is ReShelter, a reclaimed timber pavilion in the forest near Hvalsø, Denmark, built in winter and spring 2026 while Batra was a visiting scholar at CITA.
Builders narrate what they changed and why while recording on a smartphone or Meta Ray-Ban glasses. Pentimento extracts frames at 4 fps, estimates camera poses with COLMAP and trains the splat with gsplat. ArUco marker stickers on the structure match marked squares in a glTF export of the CAD model, and the backend computes a weighted rotation, scale and translation that registers the splat to the design.
A local instance of OpenAI Whisper transcribes the narration. The transcript and three frames from each video go to the Claude API, which sorts every adaptation into one or more of five drift categories, and users can edit both the labels and the text. Video timestamps order the scenes chronologically, and the browser view renders splat and model together with three.js and World Labs' Spark.
Roumen has previously overlaid CAD models on splats to troubleshoot machinery. In architecture, he told the Cornell Chronicle, "now it's more like the whole world is off." The paper reports no user study, and the authors describe their assessment of splat generation and alignment as qualitative. The paper does not link to code.
The paper is available now on arXiv, and Cornell's announcement is here.





