search

rpng / suo_slam

Symmetry and Uncertainty-Aware Object SLAM for 6DoF Object Pose Estimation
MIT License
133 stars 12 forks source link

Increasing error after running bundle adjustment demo #4

Open lishiyu005 opened 4 months ago

lishiyu005 commented 4 months ago

Dear author,

I try to run the script suo_slam/tree/main/thirdparty/g2opy/python/examples/object_slam_demo.py. However, I find the pose error even increase after bundle adjustment. Is there something wrong?

Thank you very much.

Best regards

nmerrill67 commented 4 months ago

Hello,

I Don't think I dove too deeply into that script and only used it for debugging the object pose factors. We found however that the BA helps when running the actual system. Sorry I cannot provide any further reasoning.

lishiyu005 commented 4 months ago

I also read the paper, but do not find the evaluation of multiple view. Only in table 1, but the results of multiple-view object poses are not better than GDRnet. Did I get it wrong?

nmerrill67 commented 4 months ago

We evaluated multiview on each dataset. We are competitive GDRNet when is trains a different net for each object, but we beat it when using a single net for all objects like we do which is shown in table 1.

nmerrill67 commented 4 months ago

Also note that we did not use the pbr data that GDRNet used, which could further boost the performance.