0
$\begingroup$

Our main use case is object detection in 3d lidar point clouds i.e. data is not in RGB-D format. We are planning to use CNN for this purpose using theano. Hardware limitations are CPU: 32 GB RAM Intel 47XX 4th Gen core i7 and GPU: Nvidia quadro k1100M 2GB. Kindly help me with recommendation for architecture.

I am thinking in the lines of 27000 input neurons on basis of 30x30x30 voxel grid but can't tell in advance if this is a good option.

Additional Note: Dataset has 4500 points on average per view per point cloud

$\endgroup$
1
$\begingroup$

First, CNNs are great for image recognition, where you usually take sub sampled windows of about 80 by 80 pixels, 27,000 input neurons is too large and it will take you forever to train a CNN on that.

Furthermore, why did you choose CNN? Why don't you try some more down to earth algorithms fisrst? Like SVMs, or Logistic regressions.

4500 Data points and 27000 features seems unrealistic to me, and very prone to over fitting.

Check this first.

http://scikit-learn.org/stable/tutorial/machine_learning_map/

$\endgroup$
  • $\begingroup$ SVMs are more suitable for our use case but my team wanted to explore more into deep networks and how will these help us. DBNs solved our water scheduling problem with .1% more accuracy and lesser time as compared to shallow networks so we are wondering if such an approach might help in this case as well. How would you suggest we deal with point clouds data for input to network where it has (x,y,z,r,g,b) values for each point ? $\endgroup$ – Muhammad Umar Farooq Feb 4 '15 at 6:30
  • $\begingroup$ Remember DBNs and CNNs are not the same thing, they sound similar, but they are different architectures. Your input dimension is clearly to big, so you should segment it in small windows and pretrain the CNN with those windows first. Then you should do a sweep over each lidar image, augmenting in that way your dataset, from 4500 to way more. If yoh have a single label per lidar candidate, you probably want to pool the windows for each candidate. $\endgroup$ – Leon palafox Feb 4 '15 at 16:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.