Export
Research Overview
Borislav Antic focuses on computer vision, specifically video analysis for abnormality detection and group activity recognition. His trajectory spans from parsing video frames to identify irregular behaviors without explicit training examples, to developing sample-adaptive feature weighting techniques for visual recognition. The most recent work in the evidence addresses per-sample kernel adaptation to mitigate noise and occlusion in visual data.
Research Trajectory
Borislav Antic’s early research focused on the challenges of identifying irregular events in video data, particularly when explicit examples of such events are unavailable. In 2011, he published first-author work on this topic with Video parsing for abnormality detection. The core problem addressed in this research is that the class of all irregular objects and behaviors is effectively infinite, making it impossible to gather sufficient abnormal training samples. While the standard paradigm at the time involved searching for individual abnormal local patches independently, Antic’s approach involved parsing video frames by establishing a set of hypotheses designed to jointly explain the entire scene, thereby providing a more holistic method for anomaly detection.
By 2014, Antic’s focus expanded to the complex dynamics of multi-person interactions in video. In Learning Latent Constituents for Recognition of Group Activities in Video, he investigated how the collective activity of a group is fundamentally more than the mere sum of individual actions. Recognizing that interactions and overall group context have a crucial influence on behavior, this first-author work challenged the trend of relying solely on increasingly global representations of individual bounding boxes. Instead, Antic emphasized that activities are often defined by semi-local characteristics and their interrelations across different persons. To capture large visual variability without requiring infeasible amounts of manual annotation, the research proposed learning latent constituents. This methodology aligns closely with the broader area of Joint Optimization of Task-Driven Representations and Classifiers, where feature extraction and classification are integrated rather than treated as independent sequential stages.
In 2015, Antic addressed fundamental issues of noise and occlusion in visual recognition systems. In Per-Sample Kernel Adaptation for Visual Recognition and Grouping, he noted that partial occlusion, clutter, and excessive articulation typically affect only a subset of feature dimensions, and crucially, these corrupted dimensions vary from sample to sample. The common approach in feature selection had been to either discard entire training samples or globally down-weight the same dimensions across all samples, leading to a loss of valuable signal. Antic’s work proposed adjusting the contribution of individual feature dimensions dynamically when comparing any two samples. This contribution directly advances the research direction of Sample-Adaptive Feature Weighting, moving away from global feature importance toward localized, per-sample evaluations, and touches upon themes such as Reclaiming Latent Information from Noisy Training Data by ensuring that models do not overfit to corrupted data points.
Current Directions
The most recent papers in the provided evidence are from 2015, marking a distinct focus on adaptive feature selection and noise mitigation in visual recognition. Recent work investigates how to dynamically adjust feature weights to account for sample-specific corruptions like clutter and articulation, rather than relying on global feature elimination. This focus suggests an ongoing interest in robust machine learning techniques, particularly within the research direction of Spatially Adaptive Noise Modeling, where algorithms must account for heteroscedastic noise that varies across different parts of an image or dataset. Furthermore, the mathematical complexity of adapting kernels on a per-sample basis connects to the broader area of Difference of Convex Programming, reflecting a sustained engagement with the non-convex optimization challenges inherent in modern computer vision tasks.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake