Oktober 11, 2018 3 min to read
More about Point-Clouds and Deep Learning – A paper a day…
Category : Data Science, Deep Learning, Teaching, Use-Cases
Namaste! It is perfectly possible that I might start an article-series: „A paper a day keeps the doctor away“. Named after a potential first blog-post in that series. You see, Deep-Learning-for-a-living is not so much about training Neural Nets. I repeat that mantra like a prayer. Deep Learning is more about the whole getting-and-preparing-your-data-roundtrip. And then you do training. Both are learnable skills that I teach in all my trainings. In order to stay successful, one thing is a fact for Deep Learning, that is definitely true for all other fields: You have to stay up to date! This is the challenge. But it is a positively great one. Why? Well, learning is something that is a lot of fun!
Some thoughts about learning.
As you might know, I am a passionate learner. There is no teaching without learning. I have always been passionate. But truth be told, only in my fields of interest. Do not confuse my understanding of the term „learning“ with the activity that you do in school. That is another cup of tea.
I seem to feel that true learning comes in a couple of different stages. Have a look:
- If you want to know something, read about it.
- If you want be know it better, write about it.
- If you want to be good at it, apply it.
- If you want to master it, teach it.
This is definitely resonates with my „business“. I read, apply, write and teach. And I encourage others to do the same. It consumes quite some time and energy. Yes. But it is absolutely worth the while. Why? This is the only true way to push innovation and progress.
Progress with point-clouds.
What is my topic for today? I feel blessed that I have the opportunity to contribute to the excellent Child Growth Monitor project. A cooperation of AI-Guru with Welthungerhilfe. The project is definitely about applying.
Please let me share something with you today. Remember point-clouds? I guess you do. There is no way to get away without knowing them. Point-clouds are huge collections of points in 3D-space. Just read about Lidar and you know what I mean. Lidar is one of the major building-blocks for autonomous driving.
Applying Deep Learning to point-clouds is not yet a problem-solved. This is very exciting. There is a lot of progress. And there is still a lot of potential for growth. It is amazing to see such a field of application evolve over a small number of years.
PointCNN is very promising.
Today, I spent some time with PointCNN. Again, what is does is hidden in its name. It is convolution mapped to point-clouds. The problem with point-clouds is that they are unstructured. This poses a challenge, if you intent to apply convolutions.
You need convolutions if you want to reduce the amount of sample-points, while at the same time increasing the amount of extracted features. This works like a charm for images. In image-processing convolution down-samples your images step-by-step while extracting more and more features. This is why Convolutional Neural Nets dominate the image processing domain.
Unfortunately point-clouds do not come in an ordered fashion. They are streams of unordered points. Images on the other hand are colors that are arranged in grids. This makes the use of convolutions very easy. But how would you extract features from a couple of unordered points instead of a couple of ordered pixels? The paper about PointCNN shows it very well. You „just“ do a nearest neighbor search and a transformation on your points and find groups that you can put through a convolution. This is a long-story short.
What is the outcome? PointCNN looks like the state-of-the art of point-cloud-classification and point-cloud-segmentation. If you inspect the comparisons with other approaches in the paper with PointCNN you will notice its superiority first. After that you will notice that there is still a huge space for improvements. Exciting times! Thanks for reading!
Stay in touch.
I hope you liked the article. Why not stay in touch? You will find me at LinkedIn, XING and Facebook. Please add me if you like and feel free to like, comment and share my humble contributions to the world of AI. Thank you!
A quick about me. I am a computer scientist with a love for art, music and yoga. I am a Artificial Intelligence expert with a focus on Deep Learning. As a freelancer I offer training, mentoring and prototyping. If you are interested in working with me, let me know. My email-address is firstname.lastname@example.org - I am looking forward to talking to you!