search

ethnhe / FFB6D

[CVPR2021 Oral] FFB6D: A Full Flow Bidirectional Fusion Network for 6D Pose Estimation.
MIT License
286 stars 72 forks source link

Can not reproduce the result of ICP(97%) #55

Open CodeLHY opened 2 years ago

CodeLHY commented 2 years ago

@ethnhe This is really a nice work, however, when I tried to reproduce the result of FFB6D+ICP using the ICP code from PVN3D I can not reproduce the result, can you give me a hand?

CodeLHY commented 2 years ago

https://github.com/ethnhe/PVN3D/blob/9a5481a82af4eb89fa8e0ab10ba22b9077b4652e/pvn3d/eval_icp.py#L124 It seems that you are using the ground truth mask for ICP refinement.

ethnhe commented 2 years ago

No, as illustrated in A.3. Implementation: PVN3D+ICP in our paper, we use the post-processed predicted mask for the whole images, using the fill label function. https://github.com/ethnhe/PVN3D/blob/9a5481a82af4eb89fa8e0ab10ba22b9077b4652e/pvn3d/eval_icp.py#L429 This procedure is slow so we pre-process and store it in a data structure. We use CPU clusters to speed up.

CodeLHY commented 2 years ago

No, as illustrated in A.3. Implementation: PVN3D+ICP in our paper, we use the post-processed predicted mask for the whole images, using the fill label function. https://github.com/ethnhe/PVN3D/blob/9a5481a82af4eb89fa8e0ab10ba22b9077b4652e/pvn3d/eval_icp.py#L429 This procedure is slow so we pre-process and store it in a data structure. We use CPU clusters to speed up.

Thank you for your quick reply, and I really appreciate PVN3D and FFB6D, both of them are excellent and solid. Good night.

CodeLHY commented 2 years ago

Hi @ethnhe , I followed your instruction but still get a lower result, so I reopened this issue. Average of all object: add: 87.39481105187926 adds: 95.32809689468347 add(-s): 91.22332797027117 All object (following PoseCNN): add: 87.82876641272729 adds: 95.32953369790825 add(-s): 90.4528140615951

could you please provide the files needed for eval_icp.py?

Dovisya commented 2 months ago

Hi @ethnhe , I followed your instruction but still get a lower result, so I reopened this issue. Average of all object: add: 87.39481105187926 adds: 95.32809689468347 add(-s): 91.22332797027117 All object (following PoseCNN): add: 87.82876641272729 adds: 95.32953369790825 add(-s): 90.4528140615951

could you please provide the files needed for eval_icp.py?

I also used the icp method in pvn3d, combined with the ffb6d code. However, the results of the test on ycb data were reduced. Have you solved this problem? Hope to get your reply, we can communicate with each other