Secondsite

I deployed an automated site-mapping pipeline end to end at Secondsite. It turns live footage from crane-mounted cameras into georeferenced aerial maps, coordinated across 11 Docker services.

The pipeline first ran on an 8 GB NVIDIA Jetson on site. After diagnosing a memory ceiling that hardware could not clear, I migrated it to AWS on EC2 and EBS.

Turning camera imagery into a survey-aligned map takes more than reconstruction. Frames have to be chosen, the crane itself masked out, and the result tied to surveyed reference points before the output can be trusted.

I drove the deployment and the migration to AWS, then benchmarked 6 server configurations against processing time and map quality to find the cheapest setup that still met accuracy targets.

I replaced the hand-tuned image masking with an AI segmentation model, SAM2, built the ArUco ground-control-point subsystem and GPS-driven frame selection, and delivered the project's first survey-aligned map.

I also consolidated duplicated computer-vision code into a shared library used by both the simulation harness and the production scheduler, growing the test suite from 42 to 210 tests across a 16K-line codebase.

The benchmark cut the cost per site map from $11.80 to $2.90, 75% lower, in exchange for 40 minutes of additional runtime. I documented the cost-versus-speed curve so the trade-off can be decided again for each site.

For masking, I chose SAM2 by comparing the quality of the final maps each candidate produced rather than proxy scores.

The ground-control-point subsystem detects 9 surveyed ArUco markers and emits an OpenDroneMap ground-control-point file, which is what converts otherwise un-georeferenced reconstruction output into a survey-aligned map.

Frame selection works from 48,598 daily GPS position readings to recover the crane's rotation center. That produced the first survey-aligned map, at 1 cm per pixel with 475 of 476 images placed, and cut a 5.7-hour processing stage down to minutes.

Repeated reconstruction failures traced back to masks that were unstable from frame to frame. A per-camera voting step fixed them.

Recovering the crane's rotation center was harder than it looked: the obvious approach was 9 m off.

A silent filename mismatch had been causing NodeODM to discard every ground control point without an error. I added regression coverage for it so it cannot come back unnoticed.

This work showed how quickly desktop-scale computer-vision assumptions break on edge hardware, and that the honest measure of a component is the final map it produces, not a proxy score. Resource budgets, cost curves, and regression tests for silent failures all need to be designed into the pipeline.