charlesq34
run XYZ+normal
Hello, can you update the README.md for running XYZ+normal? Thank you so much!
+1 for doing so. I'm also trying to run the net with normals but it crashes in sample_and_group_all. Trying to fix it if possible.
I am very happy that someone is doing the same thing. What changes have you made on the basis of the source code? @lokalmatador
So far I only used the cls_ssg model to build upon. Looking at the model implementation, especially placeholder inputs, one can see that the channel size is hardcoded as 3 and does not depend on either choosing to use normals or not:
pointclouds_pl = tf.placeholder(tf.float32, shape=(batch_size, num_point, 3))
Hence, the model can only work with 3 channels; I changed that to reflect the user's choice of using normals or not.
After fixing that, the next error I got was from inside pointnet_utils.py, more specifically in line 78 of sample_and_group_all which contained another hardcoded 3; changing that to 6 resolved that issue and now I'm looking at tf_sampling.py (line 56) as the underlying TF operator throws the error that it only can handle 3 channels and not 6.
All this is done under the assumption that if selecting normals, the channel size should increase to 6.
Anyways, it'd be great if the developers of pointnet++ could help in this discussion, as in the paper they mention to also use normal information, which the code however does not seem to support properly which is kinda disappointing (yet, being a researcher myself I know how difficult it is to keep everything up to date...).
Cheers!
Yes, I have also modified the values you mentioned before, as well as the parameters in the file pointnet_util.py.
grouped_xyz = tf.reshape(xyz, (batch_size, 1, nsample, 3))
But these changes also did not work.
It is expected that the author will be able to update the detailed steps into the README.
It kind of all boils down to the tf operators they added which - depending on how you change them (3 or 6 channels) partly work or not. Let's hope they find the time.
Have a nice week-end!
I did exactly the same thing. Hope the authors have some time to revise the code.
@lokalmatador , I tried to run XYZ+normal dataset, the open source code here is in default using XYZ only. And if you want to run the XYZ+normal dataset, you may need to modify several positions including the channel issues, for which the author did not deal with it in this open source code.
@ljjyxz123 did you succeed in running the net with xyz+normal information? I tried to modify the code, but unfortunately (see above) did not succeed. Similar attempts by other people also seemed to fail.
@lokalmatador , I didn't try to run with xyz+normal datasets. I am modifying the code to apply to my own dataset now, after that, I will try that. :)
@lokalmatador , Hi, what kind of instant messager do you use, maybe we can contact using that tool for much faster communication. :)
@lokalmatador , what's your ID there? If it is not convinient to publish that publicly, you can send me message, or you can email me through [email protected] Thanks!
@ljjyxz123 thanks for letting me know. In case, we could join our efforts.
Hi, Do you have any progress about the issue? Look forward.
OK,well, maybe we face the same problem, my own data only include the x,y,z,intensity, and I don't know how to use with the author's pointnet?
-
Set FLAGS.normal = True in train files.
-
Replace l0_xyz = point_cloud l0_points = None
by l0_xyz = tf.slice(point_cloud, [0,0,0], [-1,-1,3]) l0_points = tf.slice(point_cloud, [0,0,3], [-1,-1,3])
in model files.
- Replace pointclouds_pl = tf.placeholder(tf.float32, shape=(batch_size, num_point, 3))
with pointclouds_pl = tf.placeholder(tf.float32, shape=(batch_size, num_point, 6))
in model files
- Set FLAGS.normal = True in train files.
- Replace l0_xyz = point_cloud l0_points = None
by l0_xyz = tf.slice(point_cloud, [0,0,0], [-1,-1,3]) l0_points = tf.slice(point_cloud, [0,0,3], [-1,-1,3])
in model files.
- Replace pointclouds_pl = tf.placeholder(tf.float32, shape=(batch_size, num_point, 3))
with pointclouds_pl = tf.placeholder(tf.float32, shape=(batch_size, num_point, 6))
in model files
thanks！！！！！！！I succeeded！！！！！I love u!!!!
- Set FLAGS.normal = True in train files.
- Replace l0_xyz = point_cloud l0_points = None
by l0_xyz = tf.slice(point_cloud, [0,0,0], [-1,-1,3]) l0_points = tf.slice(point_cloud, [0,0,3], [-1,-1,3])
in model files.
- Replace pointclouds_pl = tf.placeholder(tf.float32, shape=(batch_size, num_point, 3))
with pointclouds_pl = tf.placeholder(tf.float32, shape=(batch_size, num_point, 6))
in model files
This is very useful!! But there is no need to Set FLAGS.normal = True in train files. Just use --normal and that will make it True.
I have a question: How to use rgb data along with xyz?
Please help me out.
Thanks.
