search

atenpas / gpd

Detect 6-DOF grasp poses in point clouds
BSD 2-Clause "Simplified" License
598 stars 233 forks source link

the speed is much lower when using CloudIndexed or CloudSampled #100

Closed xibeisiber closed 4 years ago

xibeisiber commented 4 years ago

Hi, Thanks very much for your algorithm! Recently I am using GPD (1.5 forward, only cpu) to generate grasp pose for specific object.

First I use yolo and depth image to give a bounding box of the object, select the points in this box, and then publish the pointcloud as CloudIndexed or CloudSampled.

I found in this way gpd's much slower than using the original whole pointcloud2 (30000 points) as input. Also, I tried to only publish the points of the object as PointCloud2 message (3000 points), it's still much slower.. Is the reason that the region to generate grasp is too small?? I also tried to add the rgb info to the pointcloud in Cloudindexed, it seems the rgb info reduces the time a bit.

I compared the time as in this figure

atenpas commented 4 years ago

Yes, you should only use points that are on the object. You should not need to use all object points, just subsample them. Maybe try with 500 out of the 3000? That should be much faster.

xibeisiber commented 4 years ago

Thanks for your reply. I turn to GPD2 with caffe using caffe now.. It's much faster now.