Abstract Attention modules can be added to neural network architectures to improve performance. This work presents an extensive comparison between several efficient attention modules for image classification and object detection, in addition to proposing a novel Attention Bias module with lower computational overhead. All measured attention modules have been efficiently re-implemented, which allows an objective comparison and evaluation of the relationship between accuracy and inference time. Our measurements show that single-image inference time increases far more (5–50%) than the increase in FLOPs suggests (0.2–3%) for a limited gain in accuracy, making computation cost an important selection criterion. Despite this increase in inference time, adding an attention module can outperform a deeper baseline ResNet in both speed and accuracy. Finally, we investigate the potential of adding attention modules to pretrained networks and show that fine-tuning is possible and superior to training from scratch. The choice of the best attention module strongly depends on the specific ResNet architecture, input resolution, batch size and inference framework.

Since the introduction of the GDPR and CCPA privacy legislation, both public and private facial image datasets are increasingly scrutinized. Several datasets have been taken offline completely and some have been anonymized. However, it is unclear how anonymization impacts face detection performance. To our knowledge, this paper presents the first empirical study on the effect of image anonymization on supervised training of face detectors. We compare conventional face anonymiz-ers with three state-of-the-art Generative Adversarial Network-based (GAN) methods, by training an off-the-shelf face detector on anonymized data. Our experiments investigate the suitability of anonymization methods for maintaining face detector performance, the effect of detectors overtraining on anonymization artefacts, dataset size for training an anonymizer, and the effect of training time of anonymization GANs. A final experiment investigates the correlation between common GAN evaluation metrics and the performance of a trained face detector. Although all tested anonymization methods lower the performance of trained face detectors, faces anonymized using GANs cause far smaller performance degradation than conventional methods. As the most important finding, the best-performing GAN, DeepPrivacy, removes identifiable faces for a face detector trained on anonymized data, resulting in a modest decrease from 91.0 to 88.3 mAP. In the last few years, there have been rapid improvements in realism of GAN-generated faces. We expect that further progression in GAN research will allow the use of Deep Fake technology for privacy-preserving Safe Fakes, without any performance degradation for training face detectors.

Neural pruning is a widely-used compression technique for Deep Neural Networks (DNNs). Recent innovations in Hardware Architectures (e.g. Nvidia Ampere Sparse Tensor Core) and N:M fine-grained Sparse Neural Network algorithms (i.e. every M-weights contains N non-zero values) reveal a promising research line of neural pruning. However, the existing N:M algorithms only address the challenge of how to train N:M sparse neural networks in a uniform fashion (i.e. every layer has the same N:M sparsity) and suffer from a significant accuracy drop for high sparsity (i.e. when sparsity > 80\%). To tackle this problem, we present a novel technique -- \textbf{\textit{DominoSearch}} to find mixed N:M sparsity schemes from pre-trained dense deep neural networks to achieve higher accuracy than the uniform-sparsity scheme with equivalent complexity constraints (e.g. model size or FLOPs). For instance, for the same model size with 2.1M parameters (87.5\% sparsity), our layer-wise N:M sparse ResNet18 outperforms its uniform counterpart by 2.1\% top-1 accuracy, on the large-scale ImageNet dataset. For the same computational complexity of 227M FLOPs, our layer-wise sparse ResNet18 outperforms the uniform one by 1.3\% top-1 accuracy. Furthermore, our layer-wise fine-grained N:M sparse ResNet50 achieves 76.7\% top-1 accuracy with 5.0M parameters. {This is competitive to the results achieved by layer-wise unstructured sparsity} that is believed to be the upper-bound of Neural Network pruning with respect to the accuracy-sparsity trade-off. We believe that our work can build a strong baseline for further sparse DNN research and encourage future hardware-algorithm co-design work. Our code and models are publicly available at \url{https://github.com/NM-sparsity/DominoSearch}.

Deep learning models have replaced conventional methods for machine learning tasks. Efficient inference on edge devices with limited resources is key for broader deployment. In this work, we focus on the tool selection challenge for inference deployment. We present an extensive evaluation of the inference performance of deep learning software tools using state-of-the-art CNN architectures for multiple hardware platforms. We benchmark these hardware-software pairs for a broad range of network architectures, inference batch sizes, and floating-point precision, focusing on latency and throughput. Our results reveal interesting combinations for optimal tool selection, resulting in different optima when considering minimum latency and maximum throughput.

We present a new traffic surveillance video analysis system, focusing on building a framework with robust and generic techniques, based on both scene understanding and moving object-of-interest detection. Since traffic surveillance is widely applied, we want to design a single system that can be reused for various traffic surveillance applications. Scene understanding provides contextual information, which improves object detection and can be further used for other applications in a traffic surveillance system. Our framework consists of two main stages: Semantic Hypothesis Generation (SHG) and Context-Based Hypothesis Verification (CBHV). In the SHG stage, a semantic region labeling engine and an appearance-based detector jointly generate the visual regions with specific features or of specific interests. The regions may also contain objects of interest, either moving or static. In the CBHV stage, a cascaded verification is performed to refine the results and smooth the detection by temporal filtering. We model the context by jointly considering spatial and scale constraints and motion saliency. Our proposed framework is validated on real-life road surveillance videos, in which objects-of-interest are moving vehicles. The results of the obtained vehicle detection outperform a recent object detection algorithm, in both precision (92.7%) and recall (92.0%). The framework is both conceptually and in the applied techniques of a generic nature and can be reused in various traffic surveillance applications, that operate, e.g. on a road crossing or in a harbor.

Radar is commonly used to detect and track ships in maritime surveillance. Unfortunately the systems are costly and do not provide any visual information about the object's type. To complement the ship identity information given by a radar system, we propose a supplementary system using active visual cameras that can robustly detect and track ships in harbours. By combining a high-quality, non real-time robust object detector with a feature point tracker with low computational complexity, it is possible to track ships in real time over long intervals and large distances. In addition to controlling pan and tilt, we dynamically control camera zoom to provide a high resolution image of the tracked object over a large range of distances. The tracking system is improved by a special motion estimation model for the feature points, which also incorporates zooming of the camera. The system is robust and sustains tracking even under challenging conditions, such as multiple viewpoints, a large variety of ships and various weather conditions. During experiments, various types of ships were successfully tracked for up to 18 minutes, and over a distance of almost 1.5km in the port of Rotterdam. The proposed system is generic and can be utilized in various tracking applications, by training the detector for a different object class.

We present a tracking framework in which we learn a HOG-based object detector in the first video frame and use this detector to localize the object in subsequent frames. We contribute and improve the tracking on the three following points. First, an occlusion-handling algorithm exploits discriminative information from the detector by dividing the object bounding box into patches and comparing each patch to the object model. Second, a drift-correction technique uses descriptive information of the object by calculating the similarity between the object in the previous frame and its shifted versions in the current frame. Third, a stochastic learning algorithm updates the object detector using single object and single background samples for selected frames only. Experiments with benchmark sequences show that the proposed tracker outperforms state-of-the-art methods on several sequences and has the smallest average location error.

This paper proposes an original moving ship detection approach in video surveillance systems, especially con- centrating on occlusion problems among ships and vegetation using context information. Firstly, an over- segmentation is performed to divide and classify by SVM (Support Vector Machine) segments into water or non-water, while exploiting the context that ships move only in water. We assume that the ship motion to be characterized by motion saliency and consistency, such that each ship distinguish itself. Therefore, based on the water context model, non-water segments are merged into regions with motion similarity. Then, moving ships are detected by measuring the motion saliency of those regions. Experiments on real-life surveillance videos prove the accuracy and robustness of the proposed approach. We especially pay attention to testing in the cases of severe occlusions between ships and between ship and vegetation. The proposed algorithm outperforms, in terms of precision and recall, our earlier work and a proposal using SVM-based ship detection.

We study traffic sign detection on a challenging large-scale real-world dataset of panoramic images. The core processing is based on the Histogram of Oriented Gradients (HOG) algorithm which is extended by incorporating color information in the feature vector. The choice of the color space has a large influence on the performance, where we have found that the CIELab and YCbCr color spaces give the best results. The use of color significantly improves the detection performance. We compare the performance of a specific and HOG algorithm, and show that HOG outperforms the specific algorithm by up to tens of percents in most cases. In addition, we propose a new iterative SVM training paradigm to deal with the large variation in background appearance. This reduces memory consumption and increases utilization of background information.

Training datasets for object detection problems are typically very large and Support Vector Machine (SVM) implementations are computationally complex. As opposed to these complex techniques, we use Stochastic Gradient Descent (SGD) algorithms that use only a single new training sample in each iteration and process samples in a stream-like fashion. We have incorporated SGD optimization in an object detection framework. The object detection problem is typically highly asymmetric, because of the limited variation in object appearance, compared to the background. Incorporating SGD speeds up the optimization process significantly, requiring only a single iteration over the training set to obtain results comparable to state-of-the-art SVM techniques. SGD optimization is linearly scalable in time and the obtained speedup in computation time is two to three orders of magnitude. We show that by considering only part of the total training set, SGD converges quickly to the overall optimum.

We present and experiment with a patch-based algorithm for the purpose of object classification in video surveillance. A feature vector is calculated based on template matching of a large set of image patches, within detected regions-of-interest (ROIs, also called blobs), of moving objects. Instead of matching direct image pixels, we use Gabor-filtered versions of the input image at several scales. We present results for a new typical video surveillance dataset containing over 9,000 object images. Additionally, we show results for the PETS 2001 dataset and another dataset from literature. Because our algorithm is not invariant to the object orientation, the set was split into four subsets with different orientation. We show the improvements, resulting from taking the object orientation into account. Using 50 training samples or higher, our resulting detection rate is on the average above 95%, which improves with the orientation consideration to 98%. Because of the inherent scalability of the algorithm, an embedded system implementation is well within reach.

Although many different types of technologies for information systems have evolved over the last decades (such as databases, video systems, the Internet and mobile telecommunication), the integration of these technologies is just in its infancy and has the potential to introduce "intelligent" systems. This paper describes the novelties of a video content analysis in a surveillance system, demonstrating the benefits for fast retrieval in huge video databases