We present an extension of the SPL-BEV (Soccer Player Localization for Bird’s-Eye-View) method, introducing Decomposable Tiled SPL-BEV (DTSPL-BEV), a variant designed to enable efficient inference on embedded devices. We show specialized tiling algorithms that enable inference on a given devices with some known memory limitations, while also ensuring the same detection results as running inference on the entire image. The tiling algorithms optimise tile configuration to minimises computational cost in terms of FLOPs, given memory constraints. We also introduce several potential network architecture changes that enable performance on par with SPL-BEV while reducing the computational cost. The code is available at https://github.com/IvarPersson/SPL-BEV.

Surveillance cameras typically study the exact same scene day and night, year after year, but utilize general pedestrian detectors trained to detect pedestrians from any viewing angle. By specializing the detector to the scene it is used on, its need for computational resources can be reduced and thereby its carbon footprint and operational cost. We propose a (optionally straightened) height normalizing image transform, (S)HeNIT, to be used as a preprocessing step to a CNN detector. It removes the pixel height variations of pedestrians moving on a ground plane. Therefore, the visual variation is reduced, and the available pixels are more evenly redistributed between close and distant pedestrians. This allows the detector to be specialized to a specific scene by fine-tuning and pruning it on rendered, scene specific, synthetic data. The transform also allows for lower resolution input images to be used, making recognition more difficult, which is desirable from a privacy standpoint.

A method for calibrating a Trinocular Linear Camera Array (TLCA) for traffic surveillance applications, such as towards smart cities, is presented. A TLCA-specific parametrization guarantees that the calibration finds a model where all the cameras are on a straight line. The method uses both a chequerboard close to the camera, as well as measured 3D points far from the camera: points measured in world coordinates, as well as their corresponding 2D points found manually in the images. Superior calibration accuracy can be obtained compared to standard methods using only a single data source, largely due to the use of chequerboards, while the line constraint in the parametrization allows for joint rectification. The improved triangulation accuracy, from 8-12 cm to around 6 cm when calibrating with 30-50 points in our experiment, allowing better road user analysis. The method is demonstrated by a proof-of-concept application where a point cloud is generated from multiple disparity maps, visualizing road user detections in 3D.

In the field of applied animal behaviour, video recordings of a scene of interest are often made and then evaluated by experts. This evaluation is based on different criteria (number of animals present, an occurrence of certain interactions, the proximity between animals and so forth) and aims to filter out video sequences that contain irrelevant information. However, such task requires a tremendous amount of time and resources, making manual approach ineffective. To reduce the amount of time the experts spend on watching the uninteresting video, this study introduces an automated watchdog system that can discard some of the recorded video material based on user‐defined criteria. A pilot study on cows was made where a convolutional neural network detector was used to detect and count the number of cows in the scene as well as include distances and interactions between cows as filtering criteria. This approach removed 38% (50% for additional filter parameters) of the recordings while only losing 1% (4%) of the potentially interesting video frames.

In computer vision, many problems can be formulated as finding a low rank approximation of a given matrix. Ideally, if all elements of the measurement matrix are available, this is easily solved in the L2-norm using factorization. However, in practice this is rarely the case. Lately, this problem has been addressed using different approaches, one is to replace the rank term by the convex nuclear norm, another is to derive the convex envelope of the rank term plus a data term. In the latter case, matrices are divided into sub-matrices and the envelope is computed for each subblock individually. In this paper a new convex envelope is derived which takes all sub-matrices into account simultaneously. This leads to a simpler formulation, using only one parameter to control the trade-of between rank and data fit, for applications where one seeks low rank approximations of multiple matrices with the same rank. We show in this paper how our general framework can be used for manifold denoising of several images at once, as well as just denoising one image. Experimental comparisons show that our method achieves results similar to state-of-the-art approaches while being applicable for other problems such as linear shape model estimation.

A novel sparse coding framework with unity range codes and the possibility to produce a discriminative dictionary is presented. The framework is, in contrast to many other works, able to handle unsupervised, supervised and semi-supervised settings. Furthermore, codes are constrained to be in unity range, which is beneficial in many scenarios. The paper presents the framework and solvers used to produce dictionaries and codes. Experiments in image reconstruction and feature learning for classification highlight the benefits with the proposed framework.

The purpose of this paper is threefold. Firstly, the paper introduces the One Triangle Three Parallelograms (OTTP) sampling strategy, which can be viewed as a way to index pixels from a given shape and image. Secondly, a framework for cascaded shape regression, including the OTTP sampling, is presented. In short, this framework involves binary pixel tests for appearance features combined with shape features followed by a large linear system for each regression stage in the cascade. The proposed solution is found to produce state-of-the-art results on the task of facial landmark estimation. Thirdly, the dependence of accuracy of the landmark predictions and the placement of the mean shape within the detection box is discussed and a method to visualize it is presented.

In this work, a coordinate solver for elastic net regularized logistic regression is proposed. In particular, a method based on majorization maximization using a cubic function is derived. This to reliably and accurately optimize the objective function at each step without resorting to line search. Experiments show that the proposed solver is comparable to, or improves, state-of-the-art solvers. The proposed method is simpler, in the sense that there is no need for any line search, and can directly be used for small to large scale learning problems with elastic net regularization.

This article proposes several improvements to an adaptive fingerprint enhancement method that is based on contextual filtering. The term adaptive implies that parameters of the method are automatically adjusted based on the input fingerprint image. Five processing blocks comprise the adaptive fingerprint enhancement method, where four of these blocks are updated in our proposed system. Hence, the proposed overall system is novel. The four updated processing blocks are: 1) preprocessing; 2) global analysis; 3) local analysis; and 4) matched filtering. In the preprocessing and local analysis blocks, a nonlinear dynamic range adjustment method is used. In the global analysis and matched filtering blocks, different forms of order statistical filters are applied. These processing blocks yield an improved and new adaptive fingerprint image processing method. The performance of the updated processing blocks is presented in the evaluation part of this paper. The algorithm is evaluated toward the NIST developed NBIS software for fingerprint recognition on FVC databases.

Raman spectroscopy is a laser-based vibrational technique that can provide spectral signatures unique to a multitude of compounds. The technique is gaining widespread interest as a method for detecting hidden explosives due to its sensitivity and ease of use. In this letter, we present a computationally efficient classification scheme for accurate standoff identification of several common explosives using visible-range Raman spectroscopy. Using real measurements, we evaluate and modify a recent correlation-based approach to classify Raman spectra from various harmful and commonplace substances. The results show that the proposed approach can, at a distance of 30 m, or more, successfully classify measured Raman spectra from several explosive substances, including nitromethane, trinitrotoluene, dinitrotoluene, hydrogen peroxide, triacetone triperoxide, and ammonium nitrate.

This paper introduces the concept of a mapogram. A mapogram may be viewed as a special form of spatiogram, which is a histogram containing additional spatial information. Additionally, this paper presents theory relevant to the creation of a proposed mapogram. A similarity measure derived from the Bhattacharyya coefficient is obtained in order to make comparisons between mapograms. Examples using a mapogram are given.

The purpose of this paper is threefold: firstly, the local successive mean quantization transform features are proposed for illumination and sensor insensitive operation in object recognition. Secondly, a split up sparse network of winnows is presented to speed up the original classifier. Finally, the features and classifier are combined for the task of frontal face detection. Detection results are presented for the MIT+CMU and the BioID databases. With regard to this face detector, the receiver operation characteristics curve for the BioID database yields the best published result. The result for the CMU+MIT database is comparable to state-of-the-art face detectors. A Matlab version of the face detection algorithm can be downloaded from http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=13701& objectType=FILE.

This paper presents the successive mean quantization transform (SMQT). The transform reveals the organization or structure of the data and removes properties such as gain and bias. The transform is described and applied in speech processing and image processing. The SMQT is considered as an extra processing step for the mel frequency cepstral coefficients commonly used in speech recognition. In image processing the transform is applied in automatic image enhancement and dynamic range compression.

This paper explores the successive mean quantization transform (SMQT) for automatic enhancement of gray-scale images. The transform is in the paper presented using set theory. The image enhancement capabilities and properties of the transform are analyzed. The transform is capable to perform both a nonlinear and a shape preserving stretch of the image histogram. Experiments and comparisons to histogram equalization are conducted.