Non-Maximum Suppression (NMS) is a post-processing step in almost every visual object detector, tasked with rapidly pruning the number of overlapping detected candidate rectangular Regions-of-Interest (RoIs) and replacing them with a single, more spatially accurate detection (in pixel coordinates). The common Greedy NMS algorithm suffers from drawbacks, due to the need for careful manual tuning. In visual person detection, most NMS methods typically suffer when analyzing crowded scenes with high levels of in-between occlusions. This paper proposes a modification on a deep neural architecture for NMS, suitable for such cases and capable of efficiently cooperating with recent neural object detectors. The method approaches the NMS problem as a rescoring task, aiming to ideally assign precisely one detection per object. The proposed modification exploits the extraction of RoI representations, semantically capturing the region’s visual appearance, from information-rich feature maps computed by the detector’s intermediate layers. Experimental evaluation on two common public person detection datasets shows improved accuracy against competing methods, with acceptable inference speed.

Non-maximum suppression (NMS) is a post-processing step in almost every visual object detector. NMS aims to prune the number of overlapping detected candidate regions-of-interest (RoIs) on an image, in order to assign a single and spatially accurate detection to each object. The default NMS algorithm (GreedyNMS) is fairly simple and suffers from severe drawbacks, due to its need for manual tuning. A typical case of failure with high application relevance is pedestrian/person detection in the presence of occlusions, where GreedyNMS doesn't provide accurate results. This paper proposes an efficient deep neural architecture for NMS in the person detection scenario, by capturing relations of neighboring RoIs and aiming to ideally assign precisely one detection per person. The presented Seq2Seq-NMS architecture assumes a sequence-to-sequence formulation of the NMS problem, exploits the Multihead Scale-Dot Product Attention mechanism and jointly processes both geometric and visual properties of the input candidate RoIs. Thorough experimental evaluation on three public person detection datasets shows favourable results against competing methods, with acceptable inference runtime requirements.

Recent advances in artificial intelligence, control and sensing technologies have facilitated the development of autonomous Unmanned Aerial Vehicles (UAVs). Detecting humans from video input captured on-the-fly from UAVs is a critical task for ensuring flight safety, mostly handled with lightweight Deep Neural Networks (DNNs). However the detection of individual people in the case of dense crowds and/or distribution shifts (i.e., significant visual differences between the training and the test sets) is still very challenging. This paper presents AUTH-Persons, a new, annotated, publicly available video dataset, that consists of both real and synthetic footage, suitable for training and evaluating aerial-view person detection algorithms. The synthetic data were collected from 8 visually distinct photorealistic outdoor environments and they mostly contain scenes with crowded areas, where heavy occlusions and high person densities pose challenges to common detectors. This dataset is employed to evaluate the generalization performance of various state-of-the-art detection frameworks, by testing them on environments that are visually distinct from those they have been trained on. Finally, given that Non-Maximum Suppression (NMS) methods at the end of person detection pipelines typically suffer in crowded scenes, the performance of various NMS algorithms is also compared in AUTH-Persons.

Online knowledge distillation has been proposed as an auspicious approach for circumventing the flaws of the conventional offline distillation (i.e., complex, and computationally and memory demanding process). In this work, a novel online self-distillation method, named Multilayer Online Self-Acquired Knowledge Distillation (MOSAKD), is proposed, aiming to develop fast-to-execute and effective models that can comply with applications with memory and computational restrictions, e.g., robotics applications. The MOSAKD method is able to mine additional knowledge both from the intermediate and the output layers of a deep neural model in an online fashion. To achieve this goal, k-nn non-parametric density estimation for estimating the unknown probability distributions of the data samples in the feature space generated by any neural layer is used. This enables us to compute the soft labels that explicitly express the similarities of the data with the classes, by directly estimating the posterior class probabilities of the data samples. The experimental evaluation on four datasets, including a dataset of synthetic images, indicates the effectiveness of the MOSAKD method and the superiority over existing online distillation methods.

Recent advances have facilitated the development and popularization of Unmanned Aerial Vehicles (UAVs) that can operate semi or fully autonomously. The real-time, accurate visual detection of UAVs is crucial for various tasks and applications including surveillance (e.g., detecting UAVs flying over restricted areas such as airports) or multi-robot systems (e.g., a swarm of UAVs that need to cooperate and avoid collisions between swarm members in GPS-denied environments). The small target-to-image ratio and large similarity with other flying objects makes the visual detection of UAVs a challenging task. In addition, data distribution shifts can have a major negative impact to UAV detection frameworks, often trained on a wide variety of datasets to achieve an adequate level of robustness. As an attempt to mitigate the effect of these issues, we present a method that can generate realistic annotated video data depicting flying UAVs, using as input real background videos and 3D UAV models. The conducted experimental evaluation showed that the synthetic data are both challenging and realistic and that detectors trained on a combination of real-world and synthetic data, exhibit an improved generalization performance, achieving better precision rates when evaluated on real datasets that are visually distinct from the corresponding real training data.

This paper addresses the important problem of leader detection in racing sports videos (e.g., cycling, boating and car racing events), as his/her proper framing is a pivotal issue in racing sports cinematography, where the events have a linear spatial deployment. Over the last few years, as autonomous drone vision and cinematography emerged, new challenges appeared in drone vision. While, until recently, most computer vision methods typically addressed still camera AV footage, drone sports cinematography typically employs moving cameras. In this paper, we solve the problem of leader detection in a group of similarly moving targets in sports videos, e.g. the leader of a sports cyclist group and his/her breakaway during a cycling event. This is very useful in drone sports cinematography, as it is important that the drone camera automatically centers on such a leader. We demonstrate that the novel method described in this paper can effectively solve the problem of leader detection in sports videos.