I'm sorry I'm so late to pull request the code:) . I have implemented CoAlign's multiscale fusion part in your framework. But the Agent-Object Pose Graph part is still not very easy to integrate here, because the installation is difficult and the multi-stage training strategy breaks the neatness of your framework. I made a note in my code to let reader refer to my repo if they need to use the pose correction.
CoAlign code supports proj_first=True/False, which was verified to be correct by visualization. I have provided the benchmark points under your AP calculation as well as the link to download the checkpoint.
Thanks again for your great work! And thank you so much for bringing more attention to my work through your framework!
yifanlu0227
commented
1 year ago
Hi Runsheng!
I'm sorry I'm so late to pull request the code:) . I have implemented CoAlign's multiscale fusion part in your framework. But the Agent-Object Pose Graph part is still not very easy to integrate here, because the installation is difficult and the multi-stage training strategy breaks the neatness of your framework. I made a note in my code to let reader refer to my repo if they need to use the pose correction.
CoAlign code supports proj_first=True/False, which was verified to be correct by visualization. I have provided the benchmark points under your AP calculation as well as the link to download the checkpoint.
Thanks again for your great work! And thank you so much for bringing more attention to my work through your framework!