Cevahir Çığla

Fatih Emre Simsek, Cevahir Cigla, Koray Kayabol

Multi-object tracking (MOT) has been dominated by the use of track by detection approaches due to the success of convolutional neural networks (CNNs) on detection in the last decade. As the datasets and bench-marking sites are published, research direction has shifted towards yielding best accuracy on generic scenarios including re-identification (reID) of objects while tracking. In this study, we narrow the scope of MOT for surveillance by providing a dedicated dataset of pedestrians and focus on in-depth analyses of well performing multi-object trackers to observe the weak and strong sides of state-of-the-art (SOTA) techniques for real-world applications. For this purpose, we introduce SOMPT22 dataset; a new set for multi person tracking with annotated short videos captured from static cameras located on poles with 6-8 meters in height positioned for city surveillance. This provides a more focused and specific benchmarking of MOT for outdoor surveillance compared to public MOT datasets. We analyze MOT trackers classified as one-shot and two-stage with respect to the way of use of detection and reID networks on this new dataset. The experimental results of our new dataset indicate that SOTA is still far from high efficiency, and single-shot trackers are good candidates to unify fast execution and accuracy with competitive performance. The dataset will be available at: sompt22.github.io

Sefa Burak Okcu, Burak Oğuz Özkalaycı, Cevahir Çığla

Face recognition applications in practice are composed of two main steps: face detection and feature extraction. In a sole vision-based solution, the first step generates multiple detection for a single identity by ingesting a camera stream. A practical approach on edge devices should prioritize these detection of identities according to their conformity to recognition. In this perspective, we propose a face quality score regression by just appending a single layer to a face landmark detection network. With almost no additional cost, face quality scores are obtained by training this single layer to regress recognition scores with surveillance like augmentations. We implemented the proposed approach on edge GPUs with all face detection pipeline steps, including detection, tracking, and alignment. Comprehensive experiments show the proposed approach's efficiency through comparison with SOTA face quality regression models on different data sets and real-life scenarios.