Talha Saruhan

A production-ready native cross-platform application for labeling images for semantic or instance segmentation and image classification.

I developed the application to address the shortcomings of available tools. It is tightly integrated with my cloud-based data labeling platform
and provides an always-online workflow, thus streamlining the workflow and allowing data collection and verification to be done automatically by the cloud service.

I was inspired by tools provided by Adobe Illustrator and other graphics programs when creating a variety of pen tools and brushes for the user. The application is designed to give the user the best experience, and it was written to be robust to network problems, user errors, or hindsight.

A cloud-based data labeling platform to streamline labeling workflows and monitoring progress.

I initially created this platform as a back end for my image labeling application. However, it became a completely client-agnostic, data-driven, and modular platform. It allows project managers to keep track of the labeling process and eliminates many types of errors by acting as the authoritative source for the dataset and annotations.

Wrote a blazingly fast matrix-matrix multiplication (DGEMM) routine utilizing handwritten SIMD instructions, including AVX and FMA ISAs, cache and register blocking, and cache-aware multithreading. The performance is on par with most available BLAS packages, including Intel's MKL.

Wrote a physically-based, multiple-importance sampled CPU path tracer in modern C++17 that uses multithreading. The renderer utilizes a custom scene description model and has its own lightweight OBJ importer.

Wrote a custom thread pool designed to keep specific jobs issued to multiple threads together on the same physical core. This tool gives the programmer more control over which threads share the same L1 and L2 cache and enables them to avoid unnecessary cache misses and transfers.

Implemented a cutting-edge video action classification network for a local visual impairment aid project. This is based on the paper: Hybrid Deep Learning Framework for Video Classification, arXiv:1504.01561.