3D Printing Detector

This repository offers a solution to the challenge presented in the HackEPS2021 hackathon by the Invelon - Intech3D company. The task is to build an AI model that is able to identify different pieces created with 3D printers.

What have we done?

The Data

We are provided with eight different 3D models (.stl files) representing the eight different objects we have to identify.

Synthetic Dataset Generation

One of the main problems of this task is the absence of a large dataset with enough data quality and quantity. This idea is one of the core aspects of most modern AI systems, automatically extracting the features from large amounts of data. Therefore, we must find a way to generate a Synthetic dataset from the given 3D objects that we can use to train a model that generalizes good enough for real-world environments.

To solve that, we have implemented an online data augmentation pipeline. The online term means that the data augmentation process is done in real-time, every time a sample is loaded in the requested batch. Given the stochasticity of the implement steps, the result is that it potentially creates an infinite number of samples to feed the model with.

Data augmentation and pre-processing steps:

  1. Projection: we project the 3D model (.stl) to a flat 2D plane from an arbitrary angle. This is similar to taking a photo of the given model from a random perspective.
  2. Offset/Scaling: we perform a random offset translation and scaling of the resulting 2D image.
  3. Background: we set a random background to the image to simulate a random real-life location of the piece to produce more realistic settings.
  4. Illumination, Normalization: we set arbitrary illumination changes to the images and finally normalize pixel values.

Computer Vision Model

We have decided to apply Deep Convolutional models with Transfer Learning techniques to overcome the data's lack of quantity and quality. We have experimented with the ResNet family model pre-trained with ImageNet , experimenting with the number of fine-tuned layers. We also replace the last fully-connected layer with a new linear one with as many neurons as the number of classes we expect.

Results & Performance Analysis

Team members:

Share this project:

Updates