Reproducibility gap on EuRoC dataset (~2x-3x higher RMSE than reported)

[eReh1926]

Hi, I am attempting to reproduce the EuRoC results from the paper using the provided Docker image and wanted to report a consistent reproducibility gap.

Setup:

• Docker image: xukuanhit/air_slam:v4
• TensorRT 8.6.1.6, CUDA 12.1
• GPU: RTX 4060
• Full pipeline (VO + map refinement)
• Evaluation: evo_ape euroc --align --correct_scale

Results (RMSE):

Sequence	Mine	Paper
MH_01_easy	0.039m	0.019m
V1_03_difficult	0.069m	0.025m
V2_02_medium	0.056m	0.018m

Things already tried:

• Rebuilt TensorRT engines from scratch
• Increased TensorRT workspace size to 4GB in all network builders (plnet.cpp, light_glue.cpp, super_point.cpp)
• Verified evaluation format (TUM vs EuRoC) — negligible difference

Is there something specific about how the engines were built or any configuration not covered in the README that might explain this gap?

[xukuanHIT]

@eReh1926 Hi, the gap is likely due to an evaluation reference mismatch. AirSLAM saves the camera poses, so we follow ORB-SLAM3 to evaluate using the ground truth of camera trajectories. You can obtain the camera ground truth here and give it a try.