Takeshi Ishita

Work Experience

Tier IV, Inc.

R&D of vehicle localization

July 2020 - Present

Mitou Program

Development of a Visual SLAM framework

April 2019 - March 2020

DeNA Co., Ltd.

April 2018 - March 2020
Part-time job

Cookpad Inc.

Design and implementation of machine learning methods for ingredient recognition from food images.

Patent

The model I proposed is granted as a patent #6306770.

Dec 2016 - Jul 2017
Part-time job

Usagee Inc.

  • Research and development of Machine Learinng & Computer Vision methods
  • Providing effective solutions to customers
May 2014 - Jan 2017
Part-time job

Education

National Institute of Technology, Tokyo College, Advanced Course
April 2017 - March 2019
Student exchange with Metropolia University of Applied Sciences
August 2017 - December 2017
National Institute of Technology, Tokyo College
April 2012 - March 2017

My works

My works are available on GitHub

1. Tadataka (under development)

This project aims to develop a Visual SLAM framework that is flexible and simple to use.

Currently implemented algorithms:

DVO (Dense Visual Odometry) [1] [2]

Estimating camera motion from RGB-D video sequence (YouTube video).

Feature Based Visual Odometry

Estimating camera motion and 3D structure from a single RGB camera (YouTube video).

2. RoadDamageDetector

_images/road-damage-1.png
Road damage detector based on SSD (Single Shot Multibox Detector).
The detailed explanation is at my Qiita blog page (in Japanese).
Trained models are published along with the source code.

What I did

  • Trained SSD(VGG16) on the RoadDamageDataset provided by Maeda et al. (2018) [3]
  • Replaced VGG16 with ResNet-101 and evaluated the performance

3. PCANet

PCANet is a neural network for image classification that trains its weights with PCA [4].
PCANet requires histogram calculation in the pooling layer. Although there was no GPU support for histogram calculation in CuPy.
I implemented the histogram calculation in CUDA and sent a pull request, which has been merged into the CuPy repository. #298

Ensemble PCANet

PCANet can train quickly. On the other hand, its representation ability is not strong.
I combined PCANet with Bagging and succeeded to increase the representation ability while keeping the training speed.
This idea is proposed to JSAI 2017.

4. SCW

Implementation of SCW (Soft Confidence-Weighted Learning) [5].
SCW is an online supervised learning algorithm which utilizes all the four salient properties:
  • Large margin training
  • Confidence weighting
  • Capability to handle non-separable data
  • Adaptive margin

Article

References

[1]Steinbrücker Frank, Jürgen Sturm, and Daniel Cremers. “Real-time visual odometry from dense RGB-D images.” Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on. IEEE, 2011.
[2]Kerl, Christian, Jürgen Sturm, and Daniel Cremers. “Robust odometry estimation for RGB-D cameras.” Robotics and Automation (ICRA), 2013 IEEE International Conference on. IEEE, 2013.
[3]Maeda, Hiroya, et al. “Road damage detection using deep neural networks with images captured through a smartphone.” arXiv preprint arXiv:1801.09454 (2018).
[4]Chan, Tsung-Han, et al. “PCANet: A simple deep learning baseline for image classification?.” IEEE transactions on image processing 24.12 (2015): 5017-5032.
[5]Wang, Jialei, Peilin Zhao, and Steven CH Hoi. “Exact soft confidence-weighted learning.” arXiv preprint arXiv:1206.4612 (2012).