New THOTH website available here

• Team
• Publications
• Research
• ERC Allegro
• Software
• Data
• Job Offers
• Seminars
• Contact
• Intranet

Nikita Dvornik, PhD student
A member of the THOTH Team, INRIA Rhone-Alpes, Grenoble, France

INRIA Rhone-Alpes 655 Avenue de l'Europe 38330 Montbonnot France

Office E-mail

: :

H115 firstname.lastname@inria.fr

About Me

I recieved the bachelor degree at the Moscow Institute of Physics and Technology (MIPT) in 2014, my master degree at INP Grenoble in 2016 and started my doctoral studies the same year. Currently, I am working through my PhD in computer vision and machine learning under supervision of Julien Mairal and Cordelia Schmid. I am developing new methods facilitating automatic scene understanding, e.g. object detection and semantic segmentation. As well, I am studying visual representations for problems where only a few annotated examples is available for training. My research interest are in scene understanding, real-time vision systems and problems with scarse data. Here, you can find my not-too-outdated-CV.

Publications

-->

• Diversity with Cooperation: Ensemble Methods for Few-Shot Classification
  Dvornik Nikita, Schmid Cordelia, Mairal Julien
  ICCV'19
  arxiv   hal   project page


• On the Importance of Visual Context for Data Augmentation in Scene Understanding
  Dvornik Nikita, Mairal Julien, Schmid Cordelia
  Submitted to TPAMI (under minor revision)
  arxiv   hal   project page


• Modeling Visual Context is Key to Augmenting Object Detection Datasets
  Dvornik Nikita, Mairal Julien, Schmid Cordelia
  ECCV'18
  arxiv   hal   project page


• BlitzNet: A Real-Time Deep Network for Scene Understanding
  Dvornik Nikita, Shmelkov Konstantin, Mairal Julien, Schmid Cordelia
  ICCV'17
  arxiv   hal   project page

Software and Data

• Context-driven Data Augmentor

  This GitHub repo implements all context-dependent augmentation routines and provides plug-and-play data augmentation module that could be used to train a scene understanding pipeline. It also contains the code to build, train and apply the explicit context CNN.

• BlitzNet: A Real-Time Deep Network for Scene Understanding

  A real-time scene understanding pipeline with state-of-the-art performance. Code for training, inference and a demo could be found at the GitHub project.

• Diversity with Cooperation: Ensemble Methods for Few-Shot Classification

  The full pipeline to train and distill ensembles for few-shot learning. It includes implementation of diversity- and cooparation-based training strategies as well as building and ensembling mean-centroid classifiers. The code could be found at the GitHub project.