Case Studies
Technical reports on AG2I's independent geospatial processing — aerotriangulation and orthorectification benchmarked against manufacturer reference solutions, automated HD-map feature extraction from imagery and HD-map vectors, and product-geometry references characterising delivered survey products.
Drone LiDAR as Geodetic Control for Legacy Mobile and Airborne Surveys
Mobile LiDAR from 2016 and 2018 and airborne LiDAR from 2009 over the TUM city campus, Munich, locked to a 2024 drone LiDAR survey used as the sole control — no ground control points, no shared sensor calibration. On control withheld from every fit, the 2016 mobile survey moves from 0.265 m to 0.041 m RMS and the 2018 repeat, delivered with up to 1.6 m of trajectory error, to 0.047 m; locked independently, the two epochs agree to 0.030 m (MAD) on stable ground.
Read the case study →Independent Drone Aerotriangulation and LiDAR Colourisation from a Raw UAV Delivery
A 962-image UAV block — nadir and oblique frames from one gimbal-mounted camera, flown with a LiDAR scanner — oriented and self-calibrated from the raw delivery alone: no published orientations, no published calibration, no ground control. Median image residual 0.232 px; camera stations within 0.036 m of the aircraft's RTK record; verified against the on-board LiDAR at a 0.020 m noise floor and against two state height benchmarks to 0.025 m. The same orientation then coloured 104.5 million LiDAR points from the raw imagery.
Read the case study →Fully Independent Aerotriangulation of a Vexcel UltraCam Dragon 4.2 Block
A full Dragon 4.2 five-head block (Graz, 350 exposures / 1,750 frames), aerotriangulated independently from the raw GNSS/IMU trajectory and event log — not the manufacturer's orientation. Check-point RMS 0.035 / 0.069 / 0.069 m (3D 0.103 m) on 10 held-out points; image RMS 1.31 px. Delivered natively into UltraMap as a drop-in substitution.
Read the case study →The AG2I Voxelite™: a Photoreal Control Surface for HD-Map Ground Truth
The AG2I Voxelite™ — an orthophotograph fused with its own elevation into a four-band (R,G,B,Z) gap-free surface — carries a known, control-based accuracy, so it serves as a certified control surface. Disparate datasets, an HD-map vector set and an uncontrolled mobile-LiDAR survey, are registered to it with OpenLineWorks and inherit its accuracy; the whole scene moves together into one conformed, certified frame. Control transfer has closed 1.18 m and 14.8 m gross misregistration onto a high-accuracy reference.
Read the case study →Facade Measurement by Stereophotogrammetry on Oblique Image Pairs
On a precisely aerotriangulated Vexcel UltraCam Dragon block, OpenStereo's multivantage viewer forms a stereo model from any two frames — including oblique pairs — so building facades are measured directly in 3D. Image σ₀ 0.64 px; the floating mark closes to the DSM at ΔZ 0.00 m.
Read the case study →The Company That Scales With Compute: Geospatial Production as Code
AG2I is a mapping company built as software — an engineering division that builds the suite, a production division that runs it, an orchestrator that plans, and a first-class verifier that gates every stage on measured accuracy. Capacity grows by adding agents and machines, not by hiring; throughput becomes a function of compute. The same manifest expresses any configuration, at any scale.
Read the case study →A Mapping Company in Software: Multi-Agent Orchestration of Geospatial Production
AG2I runs production as a coordinated team of agents — an orchestrator, deterministic specialist workers, and a first-class verifier over a task DAG and shared blackboard. A tested implementation runs a production line end to end: 6× parallel fan-out, an adaptive bundle-adjustment recovery (σ₀ 1.42 → 0.64), and a 0.25 m QA gate. LLM at the edges, deterministic core.
Read the case study →Automated Harvesting & Classification of Road-Marking Training Data from HD-Map Vectors
A road-marking training set built with zero manual annotation — HD-map vector layers projected onto a 0.076 m true-orthophoto yield 15,000 chips across 26 classes; a four-block CNN attains top-1 58.6 % and 56–95 % on whole road scenes. A controlled ablation isolates chirality: a 4×4 spatial head raises top-1 49.7 % → 58.6 % and right-arrow recall 28 % → 65 %. The OpenFeatureX harvester + FeatureCNN.
Read the case study →Independent AT & Orthorectification of UltraCam Condor 4.1 Imagery
UltraCam Condor 4.1, Gleisdorf (public dataset). Image σ₀ 0.93 px vs the manufacturer's 1.87 px; fitted-control RMS 0.044 / 0.042 / 0.053 m vs 0.152 / 0.242 / 0.096 m; AT-driven orthophoto co-registering to the factory orthophoto at NCC 0.95, ≈0.05 m absolute at control.
Read the case study →Level-03 Image-Product Geometry of UltraCam Condor 4.1
UltraCam Condor 4.1, Gleisdorf (public dataset). The Level-03 PAN, RGB, CIR and 270° products share one exterior orientation per exposure and are perspective frames, not orthophotos — frame fit 2.4 px vs the world-file's 50 px (≈18 m); PAN and colour co-register to ≈0.08 m.
Read the case study →