UMDSite

• 

  Construction Site Before Pointcloud Model

• 

  Construction Site Before Pointcloud Model

• 

  Construction Site After Pointcloud Model

• 

  Construction Site Middle Pointcloud Model

• 

  Construction Site Before/Middle Layered Pointcloud Model

• 

  Table Before Pointcloud Model

• 

  Table x Chairs Pointcloud Model

• 

  Table x Chairs Layered Pointcloud Model

• 

  Depth model Pointcloud - Side

• 

  Depth model Pointcloud - Front

Inspiration

Construction progress tracking is still manual. Most existing solutions rely on LiDAR, specialized sensors, or controlled capture conditions that do not reflect real construction sites. We wanted to see if ordinary video, something already captured on nearly every site, could be turned into real, quantitative progress data using geometry alone.

What it does

UMDSite converts standard construction video into measurable progress intelligence by comparing geometry over time. It reconstructs 3D structure when possible, detects what has been added or changed between states, and falls back to depth based modeling when full reconstruction becomes unreliable. The result is a low cost, software only approach to tracking construction progress without object labeling or specialized hardware.

How we built it

We built a multi stage pipeline that adapts to capture conditions.

• Used Structure from Motion to generate sparse and dense 3D point clouds from video
  
• Aligned time separated reconstructions and detected new geometry using spatial clustering
  
• Validated the comparison engine in a controlled environment where the target remained in frame
  
• Designed a wall specific AI depth model using monocular depth estimation to track wall growth when full 3D reconstruction fails
  
• Aggregated depth changes over time to quantify construction progress
  

Challenges we ran into

Live construction footage is chaotic. Worker movement, occlusion, unstable camera motion, and inconsistent lighting severely degrade traditional 3D reconstruction. Full Structure from Motion pipelines are fragile under these conditions, forcing us to rethink how progress should be measured when ideal data is not available.

Accomplishments that we're proud of

• Successfully reconstructed and compared 3D scenes using only video
  
• Demonstrated reliable geometric change detection in controlled environments
  
• Designed a novel wall specific depth approach that works without full 3D reconstruction
  
• Built a system that adapts to real world conditions instead of requiring perfect data
  

What we learned

We learned that geometry first approaches are powerful, but only when matched to the environment. Full 3D reconstruction is not always the right tool. Task specific depth modeling can be more practical in real construction settings. Adaptability matters more than theoretical perfection.

What's next for UMDSite

Next steps include integrating dense multi view stereo, using ICP to compute precise volumetric change, improving real time capture guidance, and packaging the system into a deployable progress monitoring tool for real construction workflows.

Built With

• colab
• colmap
• dbscan
• github
• icp
• matplotlib
• midas
• open3d
• opencv
• python
• pytorch
• sfm