search

rasmushaugaard / surfemb

SurfEmb (CVPR 2022)
https://surfemb.github.io/
MIT License
79 stars 18 forks source link

Any Plan to upgrade? #36

Closed smoothumut closed 1 year ago

smoothumut commented 1 year ago

Hi Rasmus, Your deep learning model is really powerful. There is soo much work in it. I have made it work and using it. For my custom pbr dataset, it works really good, but not perfect. My question is if you have plans to improve the model to v2? It would be great if you have so Thanks in advance Best Regards

rasmushaugaard commented 1 year ago

Hi there, We have extended SurfEmb to multi-view in EpiSurfEmb, but we don't have further plans to extend or revise SurfEmb. One of the limitations of SurfEmb and other point-estimate methods is lacking uncertainty quantification. We have recently proposed SpyroPose which estimates pose distributions. BR Rasmus

smoothumut commented 1 year ago

Hi Rasmus,

Wow, you are definetely doing a great job! I really admire all that work

Regarding EpiSurfEmb, Actually It is what I was looking for : ))

For our case, I am using surfemb and cosypose's multiview algorithm together to make the errors minimize and get a 3d reconstruction. It works good, but when surfemb couldnt get right poses, multiview algorithm makes the results worse as you described in your article. Your deep learning model is really deep to me : ) but if I understand correctly, It sounds to me that EpiSurfEmb will solve this issue by refining each pose comparing each others deep properties using epipolar geometry, instead of using just their scores and positions in each view. If it is so, it is very exiting to me. I hope you will share the code soon : ) I will be looking forward to it.

By the way I was also suprised how ground truth 2D detections changes the Pose prediction. Surfemb only use bbox of the 2D detections, not masks, Am I right? I didnt know bboxes are that important.

Thanks thanks in advance and your great work Umut

rasmushaugaard commented 1 year ago

Code for EpiSurfEmb will be released soon. Surfemb predicts the mask, so no mask is needed from the detector.

smoothumut commented 1 year ago

Hi Rasmus, is there any update on EpiSurfEmb?

rasmushaugaard commented 1 year ago

Hi Umut. I've sent you a mail with a link to the code and a trained tless model.

khanh0812 commented 8 months ago

@rasmushaugaard , could you send me the released source code for EpiSurfEmb? Many thanks

rasmushaugaard commented 8 months ago

@rasmushaugaard , could you send me the released source code for EpiSurfEmb? Many thanks

Certainly. Do you have a mail i can send it to?

khanh0812 commented 8 months ago

@rasmushaugaard my email is khanhdb@hcmute.edu.vn, many thanks for your quick response