Hi,

I was wondering why the performance of your implementation is lower than the original repo. Do you have any intuition on why this happens? I also made my own port of pointnet to keras a few months ago and it can't go beyond 82% accuracy on the validation set.

Thanks!

Could you tell the versions of the libraries used in the code ? It would be very helpful if you can provide requirements.txt file.

Hi Gary,
Thank you for sharing your code.
I was wondering at input transform and feature transform, why do we have to set it:
x = Dense(9, weights=[np.zeros([256, 9]), np.array([1, 0, 0, 0, 1, 0, 0, 0, 1]).astype(np.float32)])(x) input_T = Reshape((3, 3))(x)
f = Dense(64 * 64, weights=[np.zeros([256, 64 * 64]), np.eye(64).flatten().astype(np.float32)])(f) feature_T = Reshape((64, 64))(f)
Could you tell me why?
Thank you very much.
Minh

Hi, gary,

Awesome work, thanks for sharing.

I have made some modifies with your code, and the modified version achieved a 88.83% accuracy on the classification task.

This is the modified version:PointNet-Keras.

Hi there!

It seems like you cannot properly save the model...

model.save("model.h5") ends up in:
TypeError: can't pickle _thread.RLock objects

Any ideas? I have the feeling this has to do with the lambda layers.

Best,
Tristan

Hello Im trying to run the segmentation example,

but running the Seg_dataprep gives a lot of "Label File missing"
ending up with only Airplanes.h5.

my directory structure is:

Thank you :-)

Hello,

I was wondering if someone here might take a look at an issue I put in with the keras team. I think my rationale is correct, but I might be missing something. Short answer, I had to modify the Lambda/matmul layer as it didn't seem to work with multi_gpu_model. It was using the full batchsize instead of the batchsize/GPUs that I was expecting. I moved to the Dot layer with axes=2 and it seems to work.

Ticket

First, thank you for the code!
Do you have any idea of how to realize pointnet++ in keras?