Ulzhan Bissarinova

Monitoring of construction progress can be conducted at an elevated level using Earth observation data, enabling systematic tracking of the development process, alignment with the project time span, and improved budget management. Existing tracking methods exhibit an over-reliance on manual methods, characterized by labor costs, subjectivity due to human factors, and prolonged time consumption. This paper introduces the Construction Site Detection and Segmentation (CSDS) tool, which is based on 16 deep learning models for automatic remote detection and segmentation of construction sites, construction stages, and their components from highresolution satellite images. The core component of the tool, the CSDS dataset, encompasses annotated remote scenes capturing five distinct construction stages from early excavation to completed states and construction footings within them across 35 geographically diverse cities all over the world. Two variations of the CSDS dataset are presented: Fully Visible Objects Dataset (FVOD), which excludes partially visible objects, features 30,407 annotated construction objects, and All Objects Dataset (AOD), covering all annotations regardless of position in a scene, contains annotated 42,905 construction elements, which is currently the largest available open dataset in this area. To demonstrate the efficiency of the CSDS, state-of-the-art models YOLOv8-seg and YOLOv11-seg were evaluated across different image resolutions in medium and large sizes. In addition, shuffle-split and city-wise split validation strategies were assessed. Our experiments show that YOLOv11-seg large model on 1200 × 1200 image resolution yielded the highest overall detection accuracy, achieving mAP 50 score of 91% on FVOD and 89% on AOD during shuffle-split validation strategy.

The use of event-based cameras in computer vision is a growing research direction. However, despite the existing research on face detection using the event camera, a substantial gap persists in the availability of a large dataset featuring annotations for faces and facial landmarks on event streams, thus hampering the development of applications in this direction. In this work, we address this issue by publishing the first large and varied dataset (Faces in Event Streams) with a duration of 689 minutes for face and facial landmark detection in direct event-based camera outputs. In addition, this article presents 12 models trained on our dataset to predict bounding box and facial landmark coordinates with an mAP50 score of more than 90%. We also performed a demonstration of real-time detection with an event-based camera using our models. Our dataset and pre-trained models are publicly available for further research at https://github.com/IS2AI/faces-in-event-streams .

The use of event-based cameras in computer vision is a growing research direction. However, despite the existing research on face detection using the event camera, a substantial gap persists in the availability of a large dataset featuring annotations for faces and facial landmarks on event streams, thus hampering the development of applications in this direction. In this work, we address this issue by publishing the first large and varied dataset (Faces in Event Streams) with a duration of 689 min for face and facial landmark detection in direct event-based camera outputs. In addition, this article presents 12 models trained on our dataset to predict bounding box and facial landmark coordinates with an mAP₅₀ score of more than 90%. We also performed a demonstration of real-time detection with an event-based camera using our models.

This paper introduces a deep learning (DL) tool capable of classifying cities and revealing the features that characterize each city from a visual perspective. The study utilizes city view data captured from satellites and employs a methodology involving DL-based classification for city identification, along with an Explainable Artificial Intelligence (AI) tool to unveil definitive features of each city considered in this study. The city identification model implemented using the ResNet architecture yielded an overall accuracy of 84%, featuring 45 cities worldwide with varied geographic locations, Human Development Index (HDI), and population sizes. The portraying attributes of urban locations have been investigated using an explanatory visualization tool named Relevance Class Activation Maps (CAM). The methodology and findings presented by the current study enable decision makers, city managers, and policymakers to identify similar cities through satellite data, understand the salient features of the cities, and make decisions based on similarity patterns that can lead to effective solutions in a wide range of objectives such as urban planning, crisis management, and economic policies. Analyzing city similarities is crucial for urban development, transportation strategies, zoning, improvement of living conditions, fostering economic success, shaping social justice policies, and providing data for indices and concepts such as sustainability and smart cities for urban zones sharing similar patterns.