search

tjiiv-cprg / EPro-PnP

[CVPR 2022 Oral, Best Student Paper] EPro-PnP: Generalized End-to-End Probabilistic Perspective-n-Points for Monocular Object Pose Estimation
https://www.youtube.com/watch?v=TonBodQ6EUU
Apache License 2.0
1.11k stars 106 forks source link

bad pose when object is close to camera #61

Open tensorflowi opened 1 year ago

tensorflowi commented 1 year ago

Hi, I have trained the epro 6DoF model using my custom dataset, and the results have been generally impressive.

However, the orientation performance seems to degrade significantly when objects are positioned close to the camera, even when there is no truncation or occlusion. Interestingly, epro performs quite well when objects are at medium to far distances, regardless of truncation or occlusion.

To provide more context, the drop in performance occurs when the object occupies roughly 80% or more of the image width. The issue becomes more pronounced in situations where a close object is also truncated or occluded.

Could you provide any suggestions for improving the model's rotation accuracy for objects that nearly fill the camera view?

Thanks so much, Matthes

Lakonik commented 1 year ago

Hi Matthes! I think it could be possible that the x3d points are not properly initialized (say with very large values), so that the points get z-clipped even though the object itself is not truncated.