Multiple bboxes, positive and negative bboxes?

[ahundt]

It seems only the positive bounding boxes might be loaded, is this the case?

Also, how are the varying numbers of bboxes handled / structured?

The code for that part is split in several locations so it isn't easy to tell the design.

[ahundt]

Additional bug: There are also NaN bboxes, which aren't handled correctly.

[jinhuan-hit]

@ahundt Yeah,I also find the problem.As the issue #20 ,I have deleted the Nan valude in pcd0132cpos.txt and pcd0165cpos.txt file and run build_cgd_dataset.py.However,when I run grasp_det.py for train,the loss is nan.

[ahundt]

There are newer repositories with better results and more complete implementations available on github. It was a great starting point at the time of release, but I don't recommend using this repo/code anymore.

[jinhuan-hit]

@ahundt Thank you for your reply.

[jinhuan-hit]

@ahundt By the way,would you mind sharing the link of newer repositories with better results and more complete implementations available on github? Thanks a lot.

[ahundt]

This is mine and has a keras/tf loader, but it only supports single grasp: https://github.com/jhu-lcsr/costar_plan

I'm finding this pytorch implementation is pretty decent and runs on a real robot: https://github.com/andyzeng/visual-pushing-grasping

[jinhuan-hit]

@ahundt Thanks a lot.Yeah,I have read your paper "The CoSTAR Block Stacking Dataset: Learning with Workspace Constraints",a very interseting work about block stacking. Andyzeng uses Reinforcement Learning to learn how to pick and place.However,the environment is limited.Thank you for your kindness.

[ahundt]

Thanks!