SAM-assisted scene understanding
Exploring how promptable segmentation can turn complex construction imagery into structured regions for measurement and review.
Research track 01
ComputerVision.
Construction information is still captured manually, even when the relevant evidence is already visible. This track asks how machine perception can make field observations faster, more structured, and easier to act on.
Use cases / High level
Three ways perception can enter the construction workflow.
Lightweight edge intelligence
Designing compact YOLO and 1D CNN pipelines that move perception closer to the field—where latency, portability, and clarity matter.
Material digitization
Using reflection patterns and structured signal profiles to investigate how visual observations can support glass identification.
System logic
From an image to an engineering decision.
01
Observe
Capture a physical condition in the field.
→02
Segment
Identify relevant regions and visual cues.
→03
Infer
Translate pixels or signals into structured data.
→04
Integrate
Return the result to a usable app workflow.
Selected systems
Research translated into working model pipelines.
SAM research environment
A modular environment connecting shared measurement logic, a GPU-backed API, a local comparison interface, and an iPad frontend. The emphasis is on testing segmentation as part of a complete engineering workflow—not as an isolated model.
Window and glass-area detection
A specialized segmentation pipeline for identifying window and glass regions. Training, evaluation, export, and app integration are separated so model experiments remain reproducible while the mobile application stays lightweight.
Compact glass-analysis stack
A lightweight YOLO detector locates reflection cues; a structured RGB and mask profile then feeds a compact 1D convolutional model. The architecture is designed for portable inference and clear interfaces between perception stages.