Junya Morioka

profile image

Junya Morioka

Tokyo, Japan

Ph.D. in Science

Interest

Computer Vision, NLP, Web

Contact

mjun __at__ mjunya.com

Skill

Python PyTorch FastAPI HuggingFace
JavaScript React Svelte Astro
C++ ONNX TensorRT SIMD(AVX2, AVX-512)
Ruby Sinatra
PHP CodeIgniter
Swift
Docker
AWS
GitHub Actions
Blender

Education

Work Experience

Publication

Award

Open Source

Writing

Current Projects

Search System Development using NLP Embedding Models

I am developing a search system using embedding models with BERT and Faiss from natural language processing. The server side uses FastAPI, and the infrastructure uses both AWS and on-premises environments. The search target is 60TB of PDF documents, for which high-speed search is required. I am creating HuggingFace Datasets, training tokenizers and models, and building custom models.
huggingface

Past Projects

3d room reconstruction

3D Room Structure Reconstruction using a Monocular Camera

In recent years, the spread of VR devices and advances in technology have increased demand for reflecting real space in virtual space. In particular, technologies for reflecting indoor environments into virtual space are used in various fields such as architecture, interior design, and robotics. Reflecting real space in virtual space requires estimating the 3D structure of real space, but this typically requires expensive equipment and faces challenges in terms of computational cost and ease of measurement. In this research, I limit the estimation target to indoor environments and propose a 3D structure estimation method using a single image. While existing methods generally use depth estimation, this research proposes a novel approach based on semantic segmentation and cuboid fitting. Experiments using 3D CG show that the proposed method achieves accuracy exceeding existing methods and delivers competitive results.

6DoF Pose Estimation of Objects using Deep Learning Models

This research aims to devise a high-precision 3D object pose estimation method. High-precision 3D object pose estimation involves various challenges, including datasets for training, handling occlusion, and camera calibration. This research pursues everything from dataset creation to practical application of the models.
6DoF pose_estimation
spic

SPIC Inc. Internal System Development

I am building internal systems for SPIC Inc., which develops and sells health-related products including Lypo-C, beauty-related products, and original brand products.
qvsbase

Video Streaming Service Development

Developed a video streaming service for iOS/Android. The server side used Ruby's Sinatra, and AWS was used for the infrastructure. Firebase Authentication was used for authentication. I was responsible for backend API development, Firebase Authentication implementation, iOS app implementation with SwiftUI, and building CI/CD using Xcode Cloud and GitHub. In particular, I migrated the iOS app from UIKit to SwiftUI entirely from scratch by myself. The video streaming service has a multi-tenant architecture, enabling genre-specific streaming services to be launched easily by simply specifying the app design and related settings.

Small Object Detection using Deep Learning

Various methods for object detection based on deep learning have been proposed, but there is still significant room for improvement in the accuracy of small object detection. This research examines which processing is important for improving the detection accuracy of small objects, and proposes a novel method to improve small object detection accuracy based on the findings.
small object detection