Projects

R&D project on reference-conditioned video generation for object replacement and video inpainting. The goal is to replace a masked object in a video using a reference image while preserving temporal consistency and the original scene.

Evaluated modern video generation and inpainting models, including Wan and VACE. Investigated conditioning strategies using masks and reference images. Designed and compared pipelines for temporally consistent object replacement. Led a small applied AI team: defined experiments, reviewed results, and selected promising approaches.

Engineering-oriented research seminar on computer vision, from classical methods to modern deep learning and generative vision. The syllabus covers image analysis and linear classifiers (k-NN, SVM), optimization/backprop, CNNs and Vision Transformers, work with real data, metric learning, generative models, segmentation, object detection, video processing, and production deployment/performance including 3D topics. Hands-on sessions use PyTorch and modern CV tooling.

• Designed and maintained a coherent end-to-end CV curriculum that connects fundamentals → deep architectures → applied tasks, so students can build working baselines.
• Developed practical labs across core pipelines: detection, segmentation, tracking, and modern generative vision; updated materials annually and curated supporting readings.
• Structured assessment around research-style work

Research project applying deep learning to quantitative analysis of Atomic Force Microscopy (AFM) data. We developed a two-stage pipeline based on U-Net CNNs for detecting globular objects and estimating their heights on non-planar surfaces. The model achieved ~0.55 nm RMSE (R² = 0.978) and enables automated analysis of hundreds of objects in under a minute on CPU.

My contribution: development of the AI part of the project, including neural network architecture and code, training pipeline, experiment design, model training and evaluation.

Publication: E.V. Dubrovin, A.A. Ganichev, Measurement(Q1), 290 (2026), 123105.

When students can generate code with LLMs, “submit & grade” stops proving authorship. HW Checker shifts assessment to understanding: per-submission, code-specific, timed multiple-choice quizzes delivered via unique links, with detailed attempt telemetry and review tools.

• Designed an LLM-resilient assessment workflow: check understanding instead of only checking output.
• Supports timed, randomized quizzes; logs correctness + time spent; auto-records unanswered items on timeout.
• Integrates with an upstream LLM question generator (questions produced per student submission and imported as JSON).

"The purpose of the course is to provide graduate students of various faculties of Lomonosov Moscow State University with basic knowledge of programming and mathematics with the opportunity to use artificial neural network methods for big data analysis in their scientific research."

• Developed a course on deep learning and the PyTorch framework
• Created a set of Jupyter notebooks with course materials and assignments
• Advised and assisted graduate students with machine learning research
• Lectured and recorded a video on machine learning

• Problem. BOS provides indirect optical measurements (temperature/density-related fields), but engineering use-cases often require velocity and turbulence-related quantities that are not directly observed.
• My role. Consulted a PhD student and supported implementation, training, and evaluation of multiple ML approaches for BOS-to-flow reconstruction—testing baseline CNN/MLP models and physics-constrained variants (PINN), helping to compare trade-offs in accuracy, stability, and generalization.
• What we did. Co-developed and validated a data-assimilation pipeline that integrates experimental BOS temperature fields with physics constraints to reconstruct an axisymmetric turbulent hot-air jet, enabling estimation of flow properties without relying on a specific turbulence model.
• Publication : Complete characterization of axisymmetric turbulent jet using background oriented schlieren and physics-informed neural network — Heat Transfer Research (Q2 Vol. 56, Issue 5, pp. 17–35)

Goal: build a pipeline that takes one full-body photo and outputs an animation-ready 3D human (clothed mesh + rig), with texturing explored as a research direction. My role: Project Lead — defined the technical roadmap, selected core methods/models, and supervised implementation/integration across CV + 3D stages.

• Benchmarked ICON / ECON on in-the-wild smartphone photos and picked ECON for robustness to clothing + pose, with SMPL-X regularization for hands/face.
• Implemented auto-rigging for GLTF/GLB: transferred SMPL-X skeleton + skin weights to the reconstructed mesh via KD-tree nearest-neighbor mapping, and recomputed bind transforms for correct deformation.
• Enabled animation through Blender retargeting (Rokoko, BVH mocap) and validated skeletal deformations on the generated meshes.

Researched learned texturing: analyzed limits of vertex-color projection and explored UV-based and diffusion/transformer approaches (Texformer, TEXTure) toward consistent full-body textures.

• Developed a deep learning model to detect armed persons in CCTV video feeds
• Evaluated SOTA models (ViT, CLIP, BLIP, OWL-ViT …)
• Built a submodule for predicting weapon type
• Led a team of 4 junior/middle developers
• Created a dataset of short CCTV videos from real incidents
• Reviewed research literature in computer vision More info on the project page.

Skin photo analyzer for Derma.ru network of clinics.

• Developed a model for distinguishing psoriasis from other skin diseases
• Developed a model for predicting PUVA therapy effectiveness for certain patients
• Developed the architecture of a telemedicine service for dermatologists
• Developed a data labeling web app for medical images

Bicycle detection and recognition based on video from surveillance cameras

• Built a real-time video processing pipeline that utilized deep learning techniques on GPU servers
• Developed a cascade of deep learning models to detect and describe images of bicycles from various sources
• Trained the object detector (bicycle) for low-res and night videos
• Implemented and tested an object tracking algorithm
• Built a web service to track movement of a specific bicycle on a town map
• Developed a data labeling system and created a dataset of bicycle parts (~30K)
• Optimized computer vision models for faster execution with optimization techniques such as distilling and quantization
• Created an API server for external bike recognition

Videoanalytics module detecting the state of the door: open or closed. The module had to work on doors of different types, including glass and sliding doors, as well as to react to half-open doors and ignore people and other objects partially overlapping the doorway.

To solve the problem we used a monocular MiDaS depth detector. It was used to analyze the depth changes on the edges of the doorway.

One of the largest Russian off-road sites. Commercially successful project, more than 1,500 visitors per day (2015). In addition to the site itself, its own CMS, CRM, stock parser, and report generation mechanism were created. Integration with the API of IP telephony service providers and SMS mailings was completed.

Boundary recognition of certain document types: passports, ID cards, tickets …

It was required to define clear document boundaries, assuming that each boundary can be defined by a straight line. That is, the document is described by 4 points instead of 2, as with normal detection. The problem was solved by segmenting the photo and then post-processing it. The model was ported from PyTorch to C++ to run in the product environment.

Two-way data exchange with GIS of housing and communal services in synchronous and asynchronous modes. Connecting to third-party services, importing meter readings. Integration with FIAS base. The code is closed-source. Fragments on request.

Participated in the development of a universal Logo-based project environment for elementary and preschool education. This software has been widely used in Russian and Moscow schools.

• Development of components in Ninja and C++
• Development under MS Windows
• Development of networking protocols