Multilevel Data Enhancement Module For Image Quality Improvement in Group Activity Recognition

Abstract

The goal of Group Activity Recognition (GAR) is to detect activities carried out by groups based on visual information. GAR is applied in many fields such as surveillance, crowd observation, sports analysis, and autonomous systems. Image quality has a significant impact on visual features. States of pre-processing methods are often treated only as a detail of practical implementation of the GAR system, depriving the pre-processing stage of a scientific definition. This paper presents a Multi-Level Data Enhancement Module (MDEM) defined as a compact and repeatable method of the pre-processing that consists of four stages: stochastic data augmentation, local contrast boosting, noise suppression, and channel-wise normalization. The proposed module is defined as a composite transformation performed at the pixel level. MDEM is tested on frames from the Collective Activity Dataset and Volleyball dataset with different characteristics including light contrast changes, camera movements, blurring, and image compression. The qualitative and quantitative analysis of the intermediate processing stages prove that better local contrast can be achieved while maintaining better uniform representation through noise suppression. The new method developed has a clear set of procedures for preprocessing that can be linked to different neural network approaches for GAR. Future research will concern quantitative ablation studies, temporal consistency in video Processing, and data-driven enhancement methods.