Vasileios Belagiannis

4Machine learning papers tracked by Lacuna

1 other author named Vasileios Belagiannis: Vasileios Belagiannis (Friedrich-Alexander-Universität Erlangen-Nürnberg · 105 papers)

Is this you?

Research Overview

Vasileios Belagiannis focuses on computer vision, specifically articulated human pose estimation, multi-view 3D reconstruction, and robust optimization for deep regression. His trajectory progresses from real-time object tracking to complex, multi-person 3D pose estimation across multiple camera views. The most recent work in the evidence addresses the integration of robust loss functions into convolutional neural networks to handle outliers in regression tasks.

Research Trajectory

The early work of Vasileios Belagiannis focused on improving the reliability of real-time visual tracking systems. In classical tracking, methods relying on bounding boxes often inadvertently capture background information, which accumulates over time and causes the tracker to drift. As the first author of Segmentation Based Particle Filtering for Real-Time 2D Object Tracking (2012), Belagiannis addressed this limitation by integrating segmentation into the particle filtering process. By computing features based on segmented target regions rather than rigid bounding boxes, the approach successfully mitigated drift when tracking arbitrary objects undergoing significant scale and appearance changes.

Following this work in object tracking, Belagiannis shifted his focus toward the highly complex domain of human pose estimation, specifically tackling the challenges of multi-person, multi-view environments. This transition aligned with the research direction Navigating High Dimensional Search Spaces for Articulated Motion. Estimating the 3D pose of multiple humans simultaneously introduces a massive state space, compounded by partial occlusions and cross-view ambiguities when individual identities are not known in advance. In 3D Pictorial Structures for Multiple Human Pose Estimation (2014), Belagiannis and his collaborators proposed a novel framework to resolve these ambiguities. The method first reduced the state space by triangulating corresponding body joints detected across pairs of camera views. To filter out false positives and incorrectly mixed body parts, the team introduced a 3D pictorial structures model that effectively parsed the triangulated hypotheses into coherent individual poses.

Later that same year, Belagiannis extended this framework to incorporate the dimension of time, contributing to the broader area of 3D Reconstruction from Uncalibrated Multi-View Video and themes such as The Detection Bottleneck in Top-Down Pose Estimation. In Multiple Human Pose Estimation with Temporally Consistent 3D Pictorial Structures (2014), he addressed the reality that human body poses change in a consistent, continuous manner over time. By integrating multi-view human tracking to identify individuals before inferring their poses, the proposed temporally consistent 3D Pictorial Structures (3DPS) model significantly improved the robustness of pose estimation in unconstrained environments.

By 2015, Belagiannis began investigating the underlying optimization mechanisms of the deep learning models that were increasingly dominating computer vision. While Convolutional Neural Networks (ConvNets) had vastly improved regression-based tasks like pose estimation and landmark localization, they were typically optimized using L2 loss, making them highly sensitive to outliers. Engaging with work in the area of Hybrid Linear Quadratic Loss Functions for Robust Optimization, Belagiannis authored Robust Optimization for Deep Regression. In this paper, he proposed replacing standard loss functions with Tukey's biweight function, a robust M-estimator. This approach allowed the network to minimize the impact of abnormal sample estimations during training, yielding a more resilient regression model for complex visual tasks.

Current Directions

Based on the most recent papers in the provided evidence, dating to 2015, Belagiannis's research investigates the intersection of deep learning architectures and robust statistical optimization. Rather than treating neural network training as a standard optimization problem, this work emphasizes the need for objective functions that can naturally tolerate the "messy," heavy-tailed data often encountered in real-world computer vision applications. By embedding robust M-estimators directly into the training pipeline of Convolutional Neural Networks, this direction seeks to improve the reliability of deep regression models in tasks where sensor noise, occlusions, or extreme outliers would otherwise degrade performance.

Profile updated

See something wrong? Report a mistake