Unable to achieve accuracy shown in Benchmark

[NudtCalculus]

Hi tsattler, The results obtained by using parameters given in README (match method is LoFTR) can not achieve the accuracy shown in the benchmark, especially at (0.25m, 2 degrees) level. Would you please help us solve it? Looking forward to your reply.

[tsattler]

Can you explain what exactly you are doing?

[NudtCalculus]

We are running meshloc on the Aachen Day-Night v1.1 dataset using parameters given in README, but the result is not so precise.

[v-pnk]

Hi @NudtCalculus , Are you trying to match the highlighted results in the Table 3 in the paper?

That is:

• database of original photos (not rendered) from Aachen v1.1 dataset at 800 px resolution
• depth maps rendered from AC14 mesh
• LoFTR with none of "triangulate", "merge_3D_points", "cluster_keypoints", "covisibility_filtering", "all_matches_ransac" adjustments turned on
• precomputed retrieval pairs with NetVLAD top 50 database images
• RANSAC with min. 10,000 and max. 100,000 iterations and 20 px reprojection error threshold

[NudtCalculus]

Thanks for your reply, I'll try again.