the result is strange

[uuu686]

hello! I used my own point cloud data. That looks a little bit problematic. I hope you can help me.

test data:
I change the load_available_input_data(), it will return pc_full and others are None. The pc_full is read from a .pcd file .
segmap, rgb, depth, cam_K, pc_full, pc_colors = load_available_input_data(p, K=K)
the whole data is like this:

Experiment 1:
I generate the point cloud from my camera and no segmentation information, the grasp result looks good.

Experiment 2:
I segment object first and then generate the point cloud,the grasp result looks bad, and very different from Experiment 1.

I want to segment object first(just like Experiment 2) and grasp the box from it's frontage. Can you tell me what should I do?

[arsalan-mousavian]

You can segment objects with https://github.com/NVlabs/UnseenObjectClustering for example.