Johan Sokrates Wind

This Kaggle contest asked to build an algorithm that accurately reconstructs particle tracks from 3D observations left in silicon detectors. CERN needs to analyze hundreds of millions of collisions per second; therefore, looking to Kaggle for solutions on how to quickly and accurately estimate the particle trajectories. More than 600 teams competed for the prize pool of $25,000.

All details can be found at Kaggle.com/c/trackml-particle-identification.

My solution was based on a physical model of particle trajectories and statistical models for the trajectory noise and maximum probability track assignments. I applied advanced data structures in C++ to make my code fast compared to most of my competitors. Additionally, I made my algorithm and code completely modular so that I could benchmark and improve each part of the code separately. This allowed me to rapidly implement new features, testing, and hyperparameter tuning, ultimately leading to a first-place solution.

This Kaggle contest aimed to have a program score well on François Chollet's Artificial General Intelligence (AGI) benchmark. François argued modern machine learning algorithms are data-hungry and brittle, so he made a benchmark to measure AI skill-acquisition on unknown tasks, with the constraint that only a handful of demonstrations are shown to learn a complex task.

More than 900 teams competed for a prize pool of $20,000. In addition to winning, I was also the only one to receive the $3,000 bonus prize for exceptional performance on the test set.

All details about the competitions can be found at Kaggle.com/c/abstraction-and-reasoning-challenge.

My solution was based on program synthesis using a powerful domain-specific language (DSL) that I implemented in C++. By applying advanced algorithms, data structures, and optimization techniques, I searched a large number of candidate programs compared to most of my competitors.

The result of the project was a user-friendly software where you could make a 3dD printed statue of your body or head using the following steps:
1. Spin once on a rotating chair in front of a Kinect v2 depth camera.
2. Inspect a visualization of the reconstructed 3D model.
3. 3D print the reconstructed model.

The software was installed as a showcase at my university.

The 3D reconstruction extracted the raw depth and color data from the Kinect using the API. It ran a custom dense, visual odometry solution written in optimized OpenCL to get the camera positions. Then it fused the depth data into a point cloud using the camera positions and pruned outlier points using some heuristics. Finally, it ran a publicly available implementation of "SSD: Smooth Signed Distance Surface Reconstruction" to convert the point cloud to a surface mesh.

The reconstructed 3D model could then easily be printed using a nearby 3D printer.