microsoft / lamar-benchmark

Source code for the ECCV 2022 paper "Benchmarking Localization and Mapping for Augmented Reality".
Creative Commons Attribution 4.0 International
385 stars 36 forks source link

Effect of pnp_error_multiplier #43

Open ayehninnkhine opened 1 year ago

ayehninnkhine commented 1 year ago

Hi,

I'd like to share my findings on the "pnp_error_multiplier" in pose estimation. I'm running an experiment with the CAB dataset using a fusion image retrieval (Netvlad+APGEM), DISK feature extraction, and lightglue matcher. Using Python open3D library, I attempted to visualize the pose estimation results of CAB phone query data (validation data). The value of "pnp_error_multiplier" has a substantial effect on the pose estimation outcome according to my findings. Pose estimation results vary depending on the value of “pnp_error_multiplier”. This scenario also happens when I run the experiments with a custom dataset.

I'm not familiar with this parameter or its impact on pose estimation. Could you kindly explain why this is happening and provide any resources that can help me learn more about it? Thank you very much.

Visualization Result - https://docs.google.com/document/d/1d9DjTCQMIn7Sf36a-cpOcsu8q-ZHK_nqcodjRF0EnYM/edit?usp=sharing

pnp_error_multiplier value (0.0005) {'Rt_thresholds': [(1, 0.1), (5, 1.0)], 'recall': [0.020202020202020204, 0.08838383838383838]}

pnp_error_multiplier value (3.0) {'Rt_thresholds': [(1, 0.1), (5, 1.0)], 'recall': [0.3484848484848485, 0.5]}