Inspiration

The point convolution (PointConv) model (Wenxuan Wu, 2018) used to classify 3D points cloud has low accuracy when predicting rotated objects. For example, PointConv can classify the chair but it will predict a rotated chair as a stair. This problem is called the rotation equivariant problem and we are proposing a method to solve this.

What it does

Our model classifies 3D points cloud objects and segments 3D rooms. In addition, our model can solve rotation equivariant problems for 3D points cloud. The model reaches higher accuracy on the rotated dataset compared to PointConv.

How we built it

We followed the set abstraction layer in PointConv to propose a new set of abstraction layers called the group set abstraction layer (GpointconvSA). In our GpointconvSA, we used the G24 rotation subgroup to act on coordinates of all the points, which made a rotation equivariant model. Then, we used Pytorch to build a model with Python, and Flask to deploy our model.

Challenges we ran into

Building model and scarcity of facilities such as GPUs and servers to train the model. We also had difficulties in deploying the model to production because of the space complexity.

Accomplishments that we're proud of

We successfully built an AI model and improved the accuracy for point convolution on rotated 3D objects on classification and segmentation on Modelnet40.

What we learned

We learned how to improve other's work from reading their research paper and implemented an AI model. We also learned to deploy our model using Flask and other frameworks. In addition, we learned how to do research in a team and have a clear-cut destination for the future.

What's next for G-Pointconv CNN for 3D points cloud

The potential of 3D data emerges day by day, and there are a lot of tools to capture 3D objects such as Lidar and 3D cameras. The applications of 3D objects are increasing in manufacture, fashion, game industry, and self-driving industry. We want to improve this model to generate real 3D objects from the 3D points cloud and utilize the potential of the 3D points cloud to contribute to the AI industry.

Share this project:

Updates