Adopting effective techniques to automatically detect and identify small drones is a very compelling need for a number of different stakeholders in both the public and private sectors. This work presents three different original approaches that competed in a grand challenge on the “Drone vs. Bird” detection problem. The goal is to detect one or more drones appearing at some time point in video sequences where birds and other distractor objects may be also present, together with motion in background or foreground. Algorithms should raise an alarm and provide a position estimate only when a drone is present, while not issuing alarms on birds, nor being confused by the rest of the scene. In particular, three original approaches based on different deep learning strategies are proposed and compared on a real-world dataset provided by a consortium of universities and research centers, under the 2020 edition of the Drone vs. Bird Detection Challenge. Results show that there is a range in difficulty among different test sequences, depending on the size and the shape visibility of the drone in the sequence, while sequences recorded by a moving camera and very distant drones are the most challenging ones. The performance comparison reveals that the different approaches perform somewhat complementary, in terms of correct detection rate, false alarm rate, and average precision.

Automated detection of objects in aerial imagery is the basis for many applications, such as search and rescue operations, activity monitoring or mapping. However, in many cases it is beneficial to employ a detector on-board of the aerial platform in order to avoid latencies, make basic decisions within the platform and save transmission bandwidth. In this work, we address the task of designing such an on-board aerial object detector, which meets certain requirements in accuracy, inference speed and power consumption. For this, we first outline a generally applicable design process for such on-board methods and then follow this process to develop our own set of models for the task. Specifically, we first optimize a baseline model with regards to accuracy while not increasing runtime. We then propose a fast detection head to significantly improve runtime at little cost in accuracy. Finally, we discuss several aspects to consider during deployment and in the runtime environment. Our resulting four models that operate at 15, 30, 60 and 90 FPS on an embedded Jetson AGX device are published for future benchmarking and comparison by the community.

Pedestrian attribute inference is a demanding problem in visual surveillance that can facilitate person retrieval, search and indexing. To exploit semantic relations between attributes, recent research treats it as a multi-label image classification task. The visual cues hinting at attributes can be strongly localized and inference of person attributes such as hair, backpack, shorts, etc., are highly dependent on the acquired view of the pedestrian. In this paper we assert this dependence in an end-to-end learning framework and show that a view-sensitive attribute inference is able to learn better attribute predictions. Our proposed model jointly predicts the coarse pose (view) of the pedestrian and learns specialized view-specific multi-label attribute predictions. We show in an extensive evaluation on three challenging datasets (PETA, RAP and WIDER) that our proposed end-to-end view-aware attribute prediction model provides competitive performance and improves on the published state-of-the-art on these datasets.

Person re-identification (re-id) is the task of matching multiple occurrences of the same person from different cameras, poses, lighting conditions, and a multitude of other factors which alter the visual appearance. Typically, this is achieved by learning either optimal features or matching metrics which are adapted to specific pairs of camera views dictated by the pairwise labelled training datasets. In this work, we formulate a deep learning based novel approach to automatic prototype-domain discovery for domain perceptive (adaptive) person re-id (rather than camera pair specific learning) for any camera views scalable to new unseen scenes without training data. We learn a separate re-id model for each of the discovered prototype-domains and during model deployment, use the person probe image to select automatically the model of the closest prototype domain. Our approach requires neither supervised nor unsupervised domain adaptation learning, i.e. no data available from the target domains. We evaluate extensively our model under realistic re-id conditions using automatically detected bounding boxes with low-resolution and partial occlusion. We show that our approach outperforms most of the state-of-the-art supervised and unsupervised methods on the latest CUHK-SYSU and PRW benchmarks.

Fine-grained vehicle classification is the task of classifying make, model, and year of a vehicle. This is a very challenging task, because vehicles of different types but similar color and viewpoint can often look much more similar than vehicles of same type but differing color and viewpoint. Vehicle make, model, and year in com- bination with vehicle color - are of importance in several applications such as vehicle search, re-identification, tracking, and traffic analysis. In this work we investigate the suitability of several recent landmark convolutional neural network (CNN) architectures, which have shown top results on large scale image classification tasks, for the task of fine-grained classification of vehicles. We compare the performance of the networks VGG16, several ResNets, Inception architectures, the recent DenseNets, and MobileNet. For classification we use the Stanford Cars-196 dataset which features 196 different types of vehicles. We investigate several aspects of CNN training, such as data augmentation and training from scratch vs. fine-tuning. Importantly, we introduce no aspects in the architectures or training process which are specific to vehicle classification. Our final model achieves a state-of-the-art classification accuracy of 94.6% outperforming all related works, even approaches which are specifically tailored for the task, e.g. by including vehicle part detections.