Nazmican Calik

Machine Learning Engineer. Sometimes I write about data related stuff.

Here are some of the projects I’ve worked on, ranging from research in 3D computer vision to practical applications of generative AI.

Research Projects

Convolution Augmented Transformers for 3D Data Processing - Master’s Thesis

PyTorch | Point Cloud Processing | Deep Learning

Developed a novel architecture combining convolutional operations with transformer self-attention mechanisms for improved 3D point cloud understanding.

Key achievements:

  • Designed a custom convolutional block that enhances self-attention for 3D data
  • Achieved improved performance on ModelNet and ShapeNet datasets
  • Demonstrated better computational efficiency compared to pure transformer approaches

Technologies: PyTorch, CUDA, Point Cloud Library (PCL), Weights & Biases

3D Object Part Segmentation using Self-supervised Learning

PyTorch | Self-supervised Learning | 3D Vision

Created a self-supervised learning framework for 3D object part segmentation, eliminating the need for expensive manual annotations.

Key achievements:

  • Implemented SimCLR-based contrastive learning for 3D point clouds
  • Adapted PointNet architecture for feature extraction
  • Achieved comparable results to supervised methods on ShapeNet dataset
  • Reduced annotation requirements by 90%

Technologies: PyTorch, PointNet, SimCLR, Open3D

Deep Neural Network Based Action Recognition with Transformers

PyTorch | Video Understanding | Transformers

Developed a transformer-based model for action recognition in video sequences, focusing on temporal modeling and efficiency.

Key achievements:

  • Designed efficient temporal attention mechanisms for video processing
  • Evaluated on multiple datasets: Jester, Something-Something v2, and Kinetics
  • Achieved state-of-the-art results on hand gesture recognition tasks
  • Reduced inference time by 40% compared to 3D CNN baselines

Technologies: PyTorch, TimeSformer, Video Processing Libraries

Hackathon Projects

Paperwork Topology Sorter

Pisano Hackathon - 1st Prize | MongoDB | Express

Developed a web application to help users organize documents based on topological dependencies.

Key achievements:

  • Built efficient graph-based sorting algorithm
  • Implemented real-time collaborative features
  • Won first place among 50+ teams

Technologies: Node.js, MongoDB, Express, React

Interested in collaborating on a project or have questions about any of these works? Get in touch!