Shulou(Shulou.com)06/01 Report--

This article mainly explains "what is the network structure of GVCNN". The explanation content in this article is simple and clear, and it is easy to learn and understand. Please follow the ideas of Xiaobian to study and learn "what is the network structure of GVCNN" together.

Network structure:

Does it look familiar? Look at the row of images input at the front end, is it very similar to MVCNN? Take a look at MVCNN's network structure:

As you may have guessed, GVCNN is an improvement on MVCNN.

What is MVCNN?

It can be the first to introduce deep learning into 3D shape recognition, as early as ICCV2015 conference published, at that time has been able to run 90.1% of the results on the ModelNet40 dataset, can be said to be a master-level network.

Subsequent methods of dealing with point clouds will be compared to it.

Its idea is actually very simple. For three-dimensional objects,'take pictures' from multiple perspectives, get 12 pictures, then code 12 VGG networks, extract features, pool 12 sets of features, and classify them.

Therefore, MVCNN's shortcomings are also obvious, the network is huge. This is clearly inconsistent with today's miniaturization trend! Such a large network, let alone deployed to a Mobile device, is a desktop computer, it is difficult to run. So, there are not many followers in this network.

The author has seen an older paper, which is to project a three-dimensional object onto a sphere. It also converts three-dimensional objects into multiple two-dimensional images for processing. The difference with MVCNN is that spherical projection can better reflect the attributes of three-dimensional objects than planar projection.

Let's look at today's protagonist GVCNN. Its improvement is that 12 pictures have been grouped and weighted.

The author considers that the 12 images in MVCNN actually have the same weight, but in practice, the contribution of 12 images to classification is high and low. By reasonable weighting, the classification accuracy can be improved naturally.

The specific operation is as shown in the figure above. Each picture gets a set of feature values, scores are obtained through FC layer, and scores are grouped, for example, three groups are divided in the figure.

Then, the original pooling operation of MVCNN is carried out within the group. The final results were obtained by weighting the groups and pooling them.

Here are the results:

It could be seen that the improvement effect was still very obvious.

Thank you for your reading. The above is the content of "What is the network structure of GVCNN?" After studying this article, I believe everyone has a deeper understanding of what is the network structure of GVCNN. The specific use situation still needs to be verified by practice. Here is, Xiaobian will push more articles related to knowledge points for everyone, welcome to pay attention!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.