Research
I'm interested in robotics, computer vision and machine learning. Currently, my main focus is on improving self-supervised learning
methods by and for robotics.
My Erdos Number: 4 [László Lovász (1) -> Rajmohan Rajaraman (2) -> Dimitrios Kanoulas (3) -> Denis Hadjivelichkov (4)]
|
|
One-Shot Transfer of Affordance Regions? AffCorrs!
Denis Hadjivelichkov, Sicelukwanda Zwane, Marc Deisenroth, Lourdes Agapito, Dimitrios Kanoulas
CoRL 2022, 2022
Given a single reference image of an object with annotated affordance regions, can we segment semantically corresponding parts within a target scene? Our unsupervised model, AffCorrs, combines the useful properties of pre-trained DINO-ViT’s image descriptors and cyclic correspondences. AffCorrs is able to find corresponding affordances both for intra- and inter-class one-shot part segmentation.
|
|
Fully Self-Supervised Class Awareness in Dense Object Descriptors
Denis Hadjivelichkov, Dimitrios Kanoulas
CoRL 2021, 2021
paper /
video /
The method introduces learning of class-aware dense object descriptors that outperforms previous techniques with more robust pixel-to-pixel matches in multi-object scenarios. An example robotic application is shown - grasping of objects in clutter based on corresponding points.
|
|
Garbage Collection and Sorting with a Mobile Manipulatorusing Deep Learning and Whole-Body Control
Jingyi Liu*, Pietro Balatti*, Kirsty Ellis*, Denis Hadjivelichkov*, Danail Stoyanov, Arash Ajoudani, and Dimitrios Kanoulas
Humanoids 2020, 2021
paper /
video /
A deep neural network GarbageNet to detect different recyclable types of garbage in the wild, experimentally tested with a Whole-Body Controlled Mobile Manipulator.
|
|
Handling Joint Limits in Mobile Manipulator Whole Body Control via Reinforcement Learning
Denis Hadjivelichkov
MSc Dissertation, 2020
As part of my MSc Dissertation, improved the performance of a state-of-the-art method for mobile manipulator whole-body control by up to 24% in given custom environments. This was achieved by thorough investigation and simplification of the handcrafted reward signal used to optimize the system.
|
|
Detection of carcinogenic weeds in aerial images using supervised learning
Denis Hadjivelichkov
Bachelor Thesis, 2019
Created a dataset containing carcinogenic weeds (Heliotropium Europaeum) in aerial imagery of farm fields. Applied and developed machine learning models for binary image classification. Developed an image processing algorithm for localization of weeds within the images. Dealt with a limited data and achieved a high recall rate.
|
Other Projects
These include coursework, side projects and unpublished research work.
|
|
Autonomous Line-Following Buggy
University of Manchester
2017-06-06
In 2017, I designed and developed an automous buggy as part of a four-member team. The buggy is able to follow a line and avoid obstacles such as walls, line gaps or slopes. The buggy implements a PID control system that corrects it’s errors in line estimation in real time. The sensors used are an array of 6x digital light sensors and 1x ultrasonic sensor. The software is embedded in a PIC microcontroller.
|
|
Robust Mobile Face Recognition System
University of Manchester
2016-09-22
I worked as a summer intern with Dr. Hujun Yin at the University of Manchester. I developed a mobile video face detection and recognition system on a Raspberry Pi. The system is controlled via speech. It uses Viola-Jones and Eigenface methods for detection and recognition. It is able to actively re-learn and add new faces.
|
|