NVlabs / BundleSDF

[CVPR 2023] BundleSDF: Neural 6-DoF Tracking and 3D Reconstruction of Unknown Objects
https://bundlesdf.github.io/
Other
1.06k stars 116 forks source link

How is the mask generated in the dataset used in the Bundle SDF project? Thank you #35

Closed paul007008 closed 1 year ago

paul007008 commented 1 year ago

How is the mask generated in the dataset used in the Bundle SDF project? Thank you

monajalal commented 1 year ago

use xmem interactive_demo.py https://github.com/hkchengrex/XMem/blob/main/docs/DEMO.md

wenbowen123 commented 1 year ago

The mask is uint8 PNG file. You can obtain it in many ways, including Photoshop, XMem, Mask-RCNN, GrabCut, SAM, etc.

paul007008 commented 1 year ago

thank you I have verified using segment anything, it works well

wang-shuaikang commented 1 year ago

使用 xmem Interactive_demo.py https://github.com/hk Chengrex/XMem/blob/main/docs/DEMO.md

The author of XMem suggested to use eval.py to generate. Based on your experience, which one is easier to use Interactive_demo.py or eval.py?

wenbowen123 commented 1 year ago

@wang-shuaikang you can use segment-anything

wang-shuaikang commented 1 year ago

@wang-shuaikang你可以使用分段任何东西

I looked at segmentation anything, it's segmenting the object out, but it doesn't generate a black and white mask, or is there another step

wenbowen123 commented 1 year ago

@wang-shuaikang You just need to replace the segmented out foreground object to white, and the rest to be black.

wang-shuaikang commented 1 year ago

Thank you very much for your reply in your busy schedule. Yes, I also just realized that after splitting everything, the object in the foreground after splitting will be replaced with white, and the rest will be black, but I have encountered a problem when splitting anything in this project In the demonstration, it can mark objects with points or boxes, but when I run his amg.py file, it can be automatically generated, but it cannot generate the corresponding mask. It will generate multiple images based on a picture. Many of the multiple masks do not conform to the object, and the corresponding image mask cannot be directly generated. Sir, what code are you using to split anything to get the desired result?

wang-shuaikang commented 1 year ago

@wang-shuaikang 您只需将分割出的前景对象替换为白色,其余的替换为黑色。

@wang-shuaikang 您只需将分割出的前景对象替换为白色,其余的替换为黑色。

Sir, I used xmem Interactive_demo.py to make a mask for each frame according to @monajalal's method, but the objects in his mask are red, do we have to use a black and white mask, the mask of the red object What is the effect of the code?

掩码是 uint8 PNG 文件。您可以通过多种方式获取它,包括 Photoshop、XMem、Mask-RCNN、GrabCut、SAM 等。

wenbowen123 commented 1 year ago

The red one should also work

wang-shuaikang commented 1 year ago

The red one should also work Thank you for your reply in spite of your busy schedule, sir, have you tried using other colors of masks, and will it affect the pose estimation?

wenbowen123 commented 1 year ago

it wont affect, they will either be runnable or not runnable at all

wang-shuaikang commented 1 year ago

它不会影响,它们要么可运行,要么根本不可运行 Sir is right, I'm sorry for my naive thoughts!!!