Barış Batuhan Topal

Domain-Adaptive Self-Supervised Pre-Training for Face & Body Detection in Drawings

> Venue: Arxiv Preprint & Date: 2022

> Barış Batuhan Topal, Deniz Yuret, Tevfik Metin Sezgin

Drawings are powerful means of pictorial abstraction and communication. Understanding diverse forms of drawings, including digital arts, cartoons, and comics, has been a major problem of interest for the computer vision and computer graphics communities. Although there are large amounts of digitized drawings from comic books and cartoons, they contain vast stylistic variations, which necessitate expensive manual labeling for training domain-specific recognizers. In this work, we show how self-supervised learning, based on a teacher-student network with a modified student network update design, can be used to build face and body detectors. Our setup allows exploiting large amounts of unlabeled data from the target domain when labels are provided for only a small subset of it. We further demonstrate that style transfer can be incorporated into our learning pipeline to bootstrap detectors using a vast amount of out-of-domain labeled images from natural images (i.e., images from the real world). Our combined architecture yields detectors with state-of-the-art (SOTA) and near-SOTA performance using minimal annotation effort.

Vertex Ordering Algorithms for Graph Coloring Problem

> Venue: SIU 2020 & Date: 2020

> Kamer Kaya, Berker Demirel, Barış Batuhan Topal, Arda Aşık, İbrahim Buğra Demir

Graph Coloring Problem (GCP) is an NP-Hard problem, where a color is chosen for each vertex in a way that the total number of colors are minimized and no neighboring vertices have the same color. In this project, we tried to build a reinforcement learning (RL) model, which is an action-reward system which helps machines to learn the given environment and model, in order to solve similar and more complex problems, to solve GCP by exploitation of graph properties. In literature, there are papers on using RL for combinatorial optimization like Learning Combinatorial Optimization over Graphs, Neural Combinatorial Optimization with RL etc. in which Travelling Salesman, Minimum Vertex Cover, Maximum Cut problems are tackled. It can be argued that applying RL to NP-Hard problems is a hot topic but our problem -GCP- has not been investigated widely in the literature.

ML Research Scientist @ Pixery

> 03.2023 - Present

- Developing state-of-the-art quality enhancement models on visual media (i.e., images and videos). You can check my work on Crisp app .

- Working on Character and Face Generation & Restoration models. You can check our work on FaceOff app .

Machine Learning Intern @ ROBSEN

> 02.2020 - 07.2020

- Studied the state-of-art object detection algorithms such as YOLO, SSD, etc., and human pose estimation projects like OpenPose and AlphaPose.

- Developed a real-time activity recognition software based on 3D visual data. The software was planned to be created in 2 phases: human skeleton extraction and activity recognition. For the first part, an external SDK is embedded, and after retrieving the skeleton data, an activity recognizer model is constructed based on one of the state-of-art models.

Software Engineering Intern @ YapayTech

> 06.2019 - 09.2019

- Contributed to the company`s online chatting platform.

- Formed main components of its server and endpoints with Express.js & Request-Promise.

- Generated user settings and payment pages in the product control panel by using React.js.

- Created a database with Knex.js & PostgreSQL that holds data of users and their sessions.

Multi-modal Emotion Recognition in Comics

> 09.2021 - 01.2022

- Adopted transformer-based approaches, trained: SqueezeBERT for purely text data, ViT for purely visual data, VisualBERT for multimodal data. Outperformed 4th place in the EmoRecCom Challenge.

Face Generation in Golden Age Comics

> 02.2021 - 06.2021

- Worked on Context-based Face Generation in Golden Age Comics (US Comics between the 1930s-1950s). The model predicts the masked face by giving consecutive comic book panels to the model with a randomly selected face masked at the last frame. To generate the embeddings, Bi-LSTM and Conv-LSTM are used to create the embeddings, then DCGAN is preferred to reconstruct the masked face from the embedding.