Generate 0 positive grasp in my own dataset

[lzhnb]

Hi, I build my own dataset following your Network trainig partial.

My data storage format is as follow:

└─ roor
   ├─ objects.txt
   ├─ input
   │    ├─ example_gt.pcd
   │    ├─ example_0.pcd
   │    ├─ ...
   │    ├─ example_19.pcd
   │    └─ ...
   └─ test
        └─ ...

I don't have the xxx.csv which contains normal for each point, and my point cloud stored in .pcd is converted from a 512*512 depth map according to camera reference. (May be the lack of normal cause this trouble)

And I run ./generate_data ../cfg/generate_data.cfg (I just change the path and num_thread in the cfg file). Then the output is as follow:

Opening HDF5 file at: /opt/ply/test/train.h5
Creating dataset <labels>: 225000 x 1
Creating dataset <images>: 225000 x 60 x 60 x 12
Opening HDF5 file at: /opt/ply/test/test.h5
Creating dataset <labels>: 75000 x 1
Creating dataset <images>: 75000 x 60 x 60 x 12
===> Generating images for object 1/30: bedroom0_0
 mesh_file_path: /opt/ply/input/bedroom0_0_gt.pcd
 normals_file_path: /opt/ply/input/bedroom0_0_gt_normals.csv
extension: pcd
Failed to find match for field 'rgba'.
Loaded point cloud with 261120 points
Loaded mesh with 261120 points.
 runtime(computeNormals): 59.3458
camera: 0, #indices: 261120, #normals: 261120
===> Processing view 1/20
extension: pcd
Failed to find match for field 'rgba'.
Loaded point cloud with 220770 points
Voxelized cloud: 102234
 runtime(computeNormals): 9.9189
camera: 0, #indices: 102234, #normals: 102234
Estimating local reference frames ...
Estimated 1000 frames in 0.0226s.
Finding hand poses ...
Found 1000 hand sets in 1.03s
====> HAND SEARCH TIME: 1.1607
Generated 1000 hand sets.
Filtering grasps outside of workspace ...
Number of grasp candidates within workspace and gripper width: 1794
neighborhoods search time: 0.3013
Created 1794 images in 0.5582s
Eval GT ...
#grasps: 1794, #positives: 0, #negatives: 1794
...

The number of positive grasps is always 0 and the loop doesn't stop. I've increased the number of sample, but it doesn't work.

I also run the ./detect_grasps ../cfg/eigen_params.cfg .. $MY_PATH_TO_PCD$, and the output is as follow:

I think your GPD algorithm makes scene. But I meet some trouble in building my own dataset, it doesn't get positive grasps.

Can you help me? Looking for your reply.

[zhoumiaoqiang]

Hi, I build my own dataset following your Network trainig partial.

My data storage format is as follow:

└─ roor
   ├─ objects.txt
   ├─ input
   │    ├─ example_gt.pcd
   │    ├─ example_0.pcd
   │    ├─ ...
   │    ├─ example_19.pcd
   │    └─ ...
   └─ test
        └─ ...

I don't have the xxx.csv which contains normal for each point, and my point cloud stored in .pcd is converted from a 512*512 depth map according to camera reference. (May be the lack of normal cause this trouble)

And I run ./generate_data ../cfg/generate_data.cfg (I just change the path and num_thread in the cfg file). Then the output is as follow:

Opening HDF5 file at: /opt/ply/test/train.h5
Creating dataset <labels>: 225000 x 1
Creating dataset <images>: 225000 x 60 x 60 x 12
Opening HDF5 file at: /opt/ply/test/test.h5
Creating dataset <labels>: 75000 x 1
Creating dataset <images>: 75000 x 60 x 60 x 12
===> Generating images for object 1/30: bedroom0_0
 mesh_file_path: /opt/ply/input/bedroom0_0_gt.pcd
 normals_file_path: /opt/ply/input/bedroom0_0_gt_normals.csv
extension: pcd
Failed to find match for field 'rgba'.
Loaded point cloud with 261120 points
Loaded mesh with 261120 points.
 runtime(computeNormals): 59.3458
camera: 0, #indices: 261120, #normals: 261120
===> Processing view 1/20
extension: pcd
Failed to find match for field 'rgba'.
Loaded point cloud with 220770 points
Voxelized cloud: 102234
 runtime(computeNormals): 9.9189
camera: 0, #indices: 102234, #normals: 102234
Estimating local reference frames ...
Estimated 1000 frames in 0.0226s.
Finding hand poses ...
Found 1000 hand sets in 1.03s
====> HAND SEARCH TIME: 1.1607
Generated 1000 hand sets.
Filtering grasps outside of workspace ...
Number of grasp candidates within workspace and gripper width: 1794
neighborhoods search time: 0.3013
Created 1794 images in 0.5582s
Eval GT ...
#grasps: 1794, #positives: 0, #negatives: 1794
...

The number of positive grasps is always 0 and the loop doesn't stop. I've increased the number of sample, but it doesn't work.

I also run the ./detect_grasps ../cfg/eigen_params.cfg .. $MY_PATH_TO_PCD$, and the output is as follow:

I think your GPD algorithm makes scene. But I meet some trouble in building my own dataset, it doesn't get positive grasps.

Can you help me? Looking for your reply.

have you solved this problem?