How can I apply the model-free approach to my own objects?

[fertiliz]

Hello, thank you for your excellent contribution. Our teacher wants me to acquire real-time pose on Hikvision's industrial entry-level depth camera to integrate with robotic arm grasping. I have successfully run the demo, but I've encountered issues with the YCBV dataset directory when trying to run the model-free setup. Could you please inform me about the data and directory structure I should prepare for the model-free version on YCBV? Additionally, could you briefly explain how I can connect to the camera to infer my own objects using the model-free setup?

[wenbowen123]

What's the error you run into when running the YCBV model-free setup?

For your own use case, you need to prepare some reference images just like the way our LINEMOD and YCBV example outlines. Specifically, you need RGBD, camera-object poses, object masks. For masks you can use tools such as SAM. For poses you can consider BundleSDF. Alternatively, you can also record a RGBD video for your novel object and use BundleSDF to directly get a 3D model to run the model-based setup, just like that github repo demo shows.

[fertiliz]

[Thanks! I had run the Neural Object Field and generate step_0001000_mesh_real_world.obj in the /ref_views_16/ob_0000001/nerf folder. The execution code that resulted in an error is：“python run_ycb_video.py --ycbv_dir /home/robot/gdrnpp_bop2022/datasets/BOP_DATASETS/ycbv --use_reconstructed_mesh 1 --ref_view_dir /home/robot/下载/FoundationPose/nef-doc/ref_views_16”

The error log were： [init()] self.h5_file: [init()] Using pretrained model from /home/robot/下载/FoundationPose/learning/training/../../weights/2023-10-28-18-33-37/model_best.pth [init()] init done Traceback (most recent call last): File "/home/robot/下载/FoundationPose/run_ycb_video.py", line 149, in run_pose_estimation() File "/home/robot/下载/FoundationPose/run_ycb_video.py", line 114, in run_pose_estimation if not reader.is_keyframe(i): File "/home/robot/下载/FoundationPose/datareader.py", line 529, in is_keyframe return (key in self.keyframe_lines) AttributeError: 'YcbVideoReader' object has no attribute 'keyframe_lines'

My ycbv directory structure is(downloads from Bop data): |ycbv |--models | | models_info.json（The files used in the last file are from here. ） | | obj_xxxx.ply | |------obj_xxxx.png |--models_eval | | models_info.json | |------obj_xxxx.ply |--models_fine | | models_info.json | | obj_xxxx.ply | |------obj_xxxx.png |--test | |--000048~000059 | | | depth（Some PNG images in there, and the same goes for the folders below.） | | | mask（Some PNG images in there, and the same goes for the folders below.） | | | mask_visib（Some PNG images in there, and the same goes for the folders below.） | | | rgb（There are some PNG images in there, and the same goes for the folders below.） | | | scene_camera.json | | | scene_gt.json | | --------- | scene_gt_info.json |--train_pbr | |--000000~000049 | | | depth | | | mask | | | mask_visib | | | rgb | | | scene_camera.json | | | scene_gt.json | | --------- | scene_gt_info.json |--train_real | |--000000~000091 | | |--depth | | |--mask | | | mask_visib | | | rgb | | | scene_camera.json | | | scene_gt.json | | --------- | scene_gt_info.json |--train_synt | |--000000~000079 | | | depth | | | mask | | | mask_visib | | | rgb | | | scene_camera.json | | | scene_gt.json | | --------- | scene_gt_info.json |--ycbv_models(The code indicates that it cannot find the models_info.json file in the directory. ) | | models | | | ----- models_info.json （copy from folder ‘models’） | --- | models_info.json （copy from folder ’models‘）

[fertiliz]

[wenbowen123]

you need to have this file https://github.com/NVlabs/FoundationPose/blob/main/datareader.py#L451 you can download it from the PoseCNN website https://rse-lab.cs.washington.edu/projects/posecnn

[ddz16]

Hello, thank you for your excellent contribution. Our teacher wants me to acquire real-time pose on Hikvision's industrial entry-level depth camera to integrate with robotic arm grasping. I have successfully run the demo, but I've encountered issues with the YCBV dataset directory when trying to run the model-free setup. Could you please inform me about the data and directory structure I should prepare for the model-free version on YCBV? Additionally, could you briefly explain how I can connect to the camera to infer my own objects using the model-free setup?

Hello, I currently have the same needs and have conducted some experiments. Do you have a contact method so we can communicate? You can find my e-mail in my personal page.

[plusgrey]

What's the error you run into when running the YCBV model-free setup?

For your own use case, you need to prepare some reference images just like the way our LINEMOD and YCBV example outlines. Specifically, you need RGBD, camera-object poses, object masks. For masks you can use tools such as SAM. For poses you can consider BundleSDF. Alternatively, you can also record a RGBD video for your novel object and use BundleSDF to directly get a 3D model to run the model-based setup, just like that github repo demo shows.

Hi Bowen,

If I am planning to use my own object (with a series of required images), do I need to retrain the geometry network and the appearance network? Since I think the implicit representation network (like NeRF and SDF) can only represent a single object.

[fertiliz]

Hello, thank you for your excellent contribution. Our teacher wants me to acquire real-time pose on Hikvision's industrial entry-level depth camera to integrate with robotic arm grasping. I have successfully run the demo, but I've encountered issues with the YCBV dataset directory when trying to run the model-free setup. Could you please inform me about the data and directory structure I should prepare for the model-free version on YCBV? Additionally, could you briefly explain how I can connect to the camera to infer my own objects using the model-free setup?

Hello, I currently have the same needs and have conducted some experiments. Do you have a contact method so we can communicate? You can find my e-mail in my personal page.

you need to have this file https://github.com/NVlabs/FoundationPose/blob/main/datareader.py#L451 you can download it from the PoseCNN website https://rse-lab.cs.washington.edu/projects/posecnn

Thanks！

[fertiliz]

Hello, thank you for your excellent contribution. Our teacher wants me to acquire real-time pose on Hikvision's industrial entry-level depth camera to integrate with robotic arm grasping. I have successfully run the demo, but I've encountered issues with the YCBV dataset directory when trying to run the model-free setup. Could you please inform me about the data and directory structure I should prepare for the model-free version on YCBV? Additionally, could you briefly explain how I can connect to the camera to infer my own objects using the model-free setup?

Hello, I currently have the same needs and have conducted some experiments. Do you have a contact method so we can communicate? You can find my e-mail in my personal page.

OK，I will contact you later!

[wenbowen123]

yes you need to retrain the neural object field.

What's the error you run into when running the YCBV model-free setup? For your own use case, you need to prepare some reference images just like the way our LINEMOD and YCBV example outlines. Specifically, you need RGBD, camera-object poses, object masks. For masks you can use tools such as SAM. For poses you can consider BundleSDF. Alternatively, you can also record a RGBD video for your novel object and use BundleSDF to directly get a 3D model to run the model-based setup, just like that github repo demo shows.

Hi Bowen,

If I am planning to use my own object (with a series of required images), do I need to retrain the geometry network and the appearance network? Since I think the implicit representation network (like NeRF and SDF) can only represent a single object.

[tony-hj]

@fertiliz Hi Fertiliz, I current have the same needs like you, acquire real-time pose on RGBD camera to integrate with robotic arm grasping. I have successfully run the demo, Now I have making my own Dataset, but some problem cant slove, Can I discuss the relevant issues with you? This is my e-mail: shishuai.echo@gmail.com