About 25% of the world’s population live in informal urban settlements containing densely packed buildings (approximately 8,000 houses per square-km) which do not lend themselves favorably to state-of-the-art satellite-based building segmentation methods due to, for example, occlusion, vegetation, shadows and low resolution. To address these challenges, we introduce a novel instance segmentation and counting approach for dense buildings. Our system first extracts a conservative set of tentative building center points using a deep network for jumpstarting a Segment Anything Model 2 (SAM2) module to produce an initial over-segmentation. Second, we use a graph neural network to refine the over-segmented regions into polygons representing accurate building masks. Experiments show that our approach achieves higher accuracy in instance segmentation and counting especially in challenging densely packed building areas in Brazil, Mexico, India, Pakistan, and Kenya, for instance.

We present a novel approach to perform instance segmentation and counting for densely packed self-similar trees using a top-view RGB image sequence. We propose a solution that leverages pixel content, shape, and self-occlusion. First, we perform an initial over-segmentation of the image sequence and aggregate structural characteristics into a contour graph with temporal information incorporated. Second, using a graph convolutional network and its inherent local messaging passing abilities, we merge adjacent tree crown patches into a final set of tree crowns. Per various studies and comparisons, our method is superior to all prior methods and results in high-accuracy instance segmentation and counting despite the trees being tightly packed. Finally, we provide various forest image sequence datasets suitable for subsequent benchmarking and evaluation captured at different altitudes and leaf conditions.

Towards effective violence detection in densely crowded scenes, we introduce two novel computationally inexpensive real time pipelines. One is for automatic crowd violence detection and another for capturing the region with highest crowd concentration. The proposed violence detection architecture uses dense Histogram of Oriented Gradients (HOG) and dense Histogram of Motion Gradients (HMG) for feature extraction and Radial Bias Function Support Vector machine (RBF SVM) for classification. We further contribute by introducing a benchmark dataset, Dense Crowd Turbulence (DCT), having 120 videos of crowd violence and 120 for non-violence. DCT achieves an accuracy of 100% when evaluated with deep learning based violence detection frameworks. The violence detection architecture achieved a near the state of art accuracy of 87.3% on DCT.

Understanding highly accurate and real-time violent actions from surveillance videos is a demanding challenge. Our primary contribution of this work is divided into two parts. Firstly, we propose a computationally efficient Bag-of-Words (BoW) pipeline along with improved accuracy of violent videos classification. The novel pipeline's feature extraction stage is implemented with densely sampled Histogram of Oriented Gradients (HOG) and Histogram of Optical Flow (HOF) descriptors rather than Space-Time Interest Point (STIP) based extraction. Secondly, in encoding stage, we propose Outlier-Resistant VLAD (OR-VLAD), a novel higher order statistics-based feature encoding, to improve the original VLAD performance. In classification, efficient Linear Support Vector Machine (LSVM) is employed. The performance of the proposed pipeline is evaluated with three popular violent action datasets. On comparison, our pipeline achieved near perfect classification accuracies over three standard video datasets, outperforming most state-of-the-art approaches and having very low number of vocabulary size compared to previous BoW Models.