Internship Topics

Prof. Helmut Prendinger

Thank you for your interest in our topics!

Objective:

The Objective of the Internship is a tangible result, such as:

· Research results that stand a chance of publication in a major venue (journal or conference), whereby the Intern assumes authorship or co-authorship

· Development or co-development of competitive software, or an intelligent system that can be demonstrated

Paper writing will be supported by Prendinger Lab.

Topic 1:
Research Area: Deep Learning

Title of Research Topic: High-Speed Object Detection and Tracking onboard a Drone

Description: We investigate methods for collision avoidance of a drone flying at high speed (>50km/h), with other drones or even birds. Our goal is to develop a Deep Learning based component for detecting and tracking an obstacle (i.e., other drones, birds) using vision sensors. The challenge is to achieve high precision and high speed in obstacle detection and tracking.

We use the AirSim simulator to generate synthetic data to train our models.

Here is an example of our current results: https://youtu.be/C22gN2WUCMk

We are working with several Deep Learning models, such as YOLOX, and Multiple Object Tracking (MOT) models, using the OpenMMLab toolbox.

We are using different kind of input sensors:

· RGB camera

· Stereo camera

Research and Development:

· Research part:

o We train our DL models using different settings of hyper parameters.

· Development part:

o We implement different DL models on NVIDIA Orin and similar hardware platforms. We use DL frameworks such as Tensorflow or PyTorch.

Key references:

· Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, Min Sun, Monocular Quasi-Dense 3D Object Tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 45, Issue 2, February 2023

https://arxiv.org/abs/2103.07351

· Jinkun Cao, Xinshuo Weng, Rawal Khirodkar, Jiangmiao Pang, Kris Kitani, Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking, 2022

https://arxiv.org/abs/2203.14360

· Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, Xinggang Wang, ByteTrack: Multi-Object Tracking by Associating Every Detection Box, 2021

https://arxiv.org/abs/2110.06864

· Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, Jian Sun. YOLOX: Exceeding YOLO Series in 2021, 2021

https://arxiv.org/abs/2107.08430

· Chao Qu, Ty Nguyen, Camillo J. Taylor, Depth Completion via Deep Basis Fitting, 2019

https://arxiv.org/abs/1912.10336

Topic 2:
Research Area: Machine Learning/Deep Learning/Cointegration

Title of Research Topic: Time Series Analysis for Bitcoin Market Prediction

Description: We aim to understand the predictability of cryptocurrencies, such as Bitcoin or Ethereum.

Our goal is to develop Machine Learning (ML) and Deep Learning (DL) models for predicting the price of Bitcoin and other cryptocurrencies for successful swing trading.

We consider several ML/DL models, such as:

· XGBoost

· Conformal Prediction

· Momentum Transformers, including Change Point Detection

We apply methods for time series forecasting, such as:

· Cointegration, Transfer Entropy, Convergent Cross-Mapping

We plan to use different kinds of input:

· Candlestick chart

· Indicators, such as Relative Strength Index (RSI)

· Etc.

Research and Development:

· Research part:

o We train our DL models using different settings of hyper parameters.

· Development part:

o We implement different DL models.

Key references:

· Hansika Hewamalage, Klaus Ackermann, Christoph Bergmeir, Forecast Evaluation for Data Scientists: Common Pitfalls and Best Practices, 2022

https://arxiv.org/abs/2203.10716

· Kieran Wood, Stephen Roberts, and Stefan Zohren, Slow Momentum with Fast Reversion: A Trading Strategy Using Deep Learning and Changepoint Detection, 2021

https://arxiv.org/abs/2105.13727

· Patrick Jaquart, David Dann, Christof Weinhardt, Short-term bitcoin market prediction via machine learning. The Journal of Finance and Data Science, Vol. 7, 2021, 45-66

https://www.sciencedirect.com/science/article/pii/S2405918821000027?via%3Dihub

· Chengyi Tu, Ying Fan, Jianing Fan, Universal Cointegration and Its Applications, iScience Vol. 19, 2019, 986-995

Topic 3:
Research Area: Deep Learning

Title of Research Topic: Transformer-based Conditional Generative Models

Description: Recent advancements in stable diffusion models have demonstrated the potential for generating high-resolution images from text queries using latent representations [3]. Further developments have expanded this approach to incorporate multiple modality queries for improved sample conditioning [1,2].

This project aims to conduct experiments with and improve the latest conditioned generative and diffusion approaches, such as those described in [1,2]. Through an analysis of these models and their intermediate results, we aim to identify opportunities for improvement. This could include increasing the diversity of generated outputs (mode collapse problem) or incorporating new constraints to enhance sample generation.

Key references:

[1] controlnet: https://arxiv.org/abs/2302.05543

https://github.com/lllyasviel/ControlNet

[2] gligen: https://arxiv.org/abs/2301.07093

https://github.com/gligen/GLIGEN

[3] Stable diffusion: https://arxiv.org/abs/2112.10752

https://github.com/CompVis/latent-diffusion

https://github.com/CompVis/stable-diffusion

[4] wavenet: https://arxiv.org/abs/1609.03499

https://github.com/vincentherrmann/pytorch-wavenet

https://openreview.net/pdf?id=rJe4ShAcF7

[5] https://arxiv.org/abs/1706.03762

E-mail: helmut @ nii [Address format is: username@nii.ac.jp]

Personal Website: http://research.nii.ac.jp/~prendinger/

LOOKING FORWARD TO HEARING FROM YOU!