Refine camera poses and 3D points for large photogrammetry problems on the GPU. It is a newer Garmo Solvers module from the October 2026 benchmark campaign, outside the 1.0.2 release. Results are checked against the original input; the measured result below states where it is ahead and where it is not.
BUILT FOR
Photogrammetry, mapping, robotics and computer-vision teams with large reconstruction workloads.
CHALLENGES → PRODUCT RESPONSE
The problem you face. What the product changes.
YOUR CHALLENGEWITH BUNDLE ADJUSTMENT
01
The problem needs the right operation
A bundle-adjustment problem in the BAL format.
Refine camera poses and 3D points for large photogrammetry problems on the GPU.
02
A result must answer the original question
An answer detached from its original constraints or input is difficult to rely on.
The suite rechecks the answer against the original input and identifies the applicable result state.
03
The comparison needs a defined scope
A capability or benchmark result is not a universal promise for a different workload.
Evaluate the selected input, resource budget, complete cost and required certificate against your baseline.
WHAT YOU GET
A defined capability. A useful result.
01
Selected solving capability
Refine camera poses and 3D points for large photogrammetry problems on the GPU.
02
Original-input result
Refined cameras and points, with the final reprojection cost of the original problem reported.
03
Common access
Run it as garmo nlls or through the garmo-mcp server; headless execution fits scripts and automated workflows.
04
Checked result
The final cost is evaluated on the original problem, so it compares directly with a baseline solver.
GARMO SOLVERS / MEASURED RESULT
Measured: 24.5 s vs 168 s
GPU Levenberg-Marquardt with an implicit Schur complement, against Ceres Solver 2.2 with 32 threads on BAL. dubrovnik-356 reaches Ceres's final cost in 24.5 s vs 168 s and ends lower (491,384 vs 493,850); final-93 in 2.2 s vs 4.7 s. Ceres is faster on small problems (under about 2 s), trafalgar-257 and venice-52.
Source: Garmo Labs benchmark campaign, October 2026. Results apply to the stated inputs, hardware and conditions; they are not universal competitor rankings.
A customer has a supported model and wants an accepted result without adopting every other module.
Illustrative customer use—not a new benchmark or completed customer deployment.
01→
Provide the original input and define the useful outcome.
02→
Run the selected module under its supported resource and interface scope.
03→
Review the returned answer, original-input checks and completion state.
FIT & ACCEPTANCE
Start with the task. Agree the useful outcome.
What you bring
A bundle-adjustment problem in the BAL format.
What a useful evaluation measures
+Accepted original-model result
+Appropriate certificate or disclosed uncertified state
+Total time, memory and operating cost
Delivery
License the selected capability, commission an implementation, or agree an operated service. Infrastructure and additional engineering are specified separately.
QUESTIONS BEFORE BUYING
A clear scope. A clear relationship.
What is the important boundary?−
The GPU path is ahead on large problems; Ceres is faster on small problems and on some reported large cases.
How is it priced?+
No new standalone price is supplied for this product; request a scoped proposal. The platform plans buy qualified workflows and capacity, not a second charge every time a compatible method is selected. Existing bundle inclusions do not expand automatically.
How do we get access?+
Available through garmo CLI and garmo-mcp, including headless execution. Product-specific inputs, runtime dependencies, budgets and result checks remain explicit.
PUT THE PRODUCT TO WORK
Tell us what needs to improve.
Bring a task, an accepted outcome and the environment in which the product needs to work.
Select the capability you need directly or include its compatible role in a qualified platform workflow. Production scope identifies release, environment, capacity and support; a product count is not an entitlement.
