This paper presents a theoretically grounded optimization framework for neural network training that integrates an Exponentially Decaying Learning Rate with Lyapunov-based stability analysis. We develop a dynamic learning rate algorithm and prove that it induces connected and stable descent paths through the loss landscape by maintaining the connectivity of super-level sets Sλ = {θ ∈ ℝn : ℒ(θ) ≥ λ}. Under the condition that the Lyapunov function V(θ) = ℒ(θ) satisfies Δ V(θ) ⋅ Δ ℒ(θ) ≥ 0, we establish that these super-level sets are not only connected but also equiconnected across epochs, providing uniform topological stability. We further derive convergence guarantees using a second-order Taylor expansion and demonstrate that our exponentially scheduled learning rate with gradient-based modulation leads to a monotonic decrease in loss. The proposed algorithm incorporates this schedule into a stability-aware update mechanism that adapts step sizes based on both curvature and energy-level geometry. This work formalizes the role of topological structure in convergence dynamics and introduces a provably stable optimization algorithm for high-dimensional, non-convex neural networks.

The generation of high-resolution boreal peatland maps through the integration of multi-source remote sensing data provides critical support for greenhouse gas inventories and environmental management. In this study, we propose GeoFusion, a pixel-wise peatland classification architecture that leverages open-access radar imagery from Sentinel-1 and optical imagery from Sentinel-2. GeoFusion combines convolutional neural networks (CNNs) for hierarchical local feature extraction with vision transformers (ViTs) for modeling long-range dependencies, enabling effective fusion of spectral–textural and structural information. The architecture also incorporates data augmentation, multi-head self-attention, and dropout regularization to enhance robustness and mitigate overfitting. Performance was assessed at two classification levels: site type and fertility level, capturing both ecological diversity and functional characteristics of complex peatland ecosystems. Comparative experiments against well-known deep learning models and hybrid Transformer-based models were conducted across three ecologically distinct regions in Finland. The obtained results demonstrate that GeoFusion outperforms other models and achieves average accuracies of 79.25% and 72.68% for fertility-level and site-type predictions,

This paper presents a theoretically grounded optimization framework for neural network training that integrates an Exponentially Decaying Learning Rate with Lyapunov-based stability analysis. We develop a dynamic learning rate algorithm and prove that it induces connected and stable descent paths through the loss landscape by maintaining the connectivity of super-level sets Sλ = {θ ∈ ℝn : ℒ(θ) ≥ λ}. Under the condition that the Lyapunov function V(θ) = ℒ(θ) satisfies Δ V(θ) ⋅ Δ ℒ(θ) ≥ 0, we establish that these super-level sets are not only connected but also equiconnected across epochs, providing uniform topological stability. We further derive convergence guarantees using a second-order Taylor expansion and demonstrate that our exponentially scheduled learning rate with gradient-based modulation leads to a monotonic decrease in loss. The proposed algorithm incorporates this schedule into a stability-aware update mechanism that adapts step sizes based on both curvature and energy-level geometry. This work formalizes the role of topological structure in convergence dynamics and introduces a provably stable optimization algorithm for high-dimensional, non-convex neural networks.

Abstract Human pose estimation has gained significant attention in recent years for its potential to revolutionize athletic performance analysis, enhance understanding of player interactions, and optimize training regimes. Deep learning models, particularly Convolutional Neural Networks (CNNs), have outperformed traditional methods in pose estimation tasks. This study addresses a gap in sports analytics by applying two popular CNN-based frameworks, YOLO and DeepLabCut, to analyze pose estimation in hurdles athletes. Videos of a single female athlete during training sessions were used, and frames were manually annotated to capture three critical foot landmarks: ankle, heel, and big toe. The results highlight YOLOv8l’s superior accuracy, achieving a Percentage of Correct Keypoints (PCK) of 79% for these landmarks, while demonstrating the feasibility of a low-cost setup for practical applications. Visual comparisons further validate the model’s effectiveness in real-world scenarios. Additionally, YOLO predictions were utilized to analyze step progression in the time domain, providing actionable insights into athletic movement. This study underscores that even modest video equipment, combined with CNN-based methods, can equip coaches with powerful tools to analyze and optimize movements and techniques, paving the way for data-driven advancements in sports performance.

Peatland classification provides valuable information for greenhouse gas inventory and biodiversity protection. In this paper, we proposed an encoder-decoder-based architecture for peatland classification that fuses two open-source satellite data, Sentinel-1 and Sentinel-2. We show the effect of fusion by comparing the multi-modal fusion architecture with unimodals which are trained only based on one input data source. We also investigate the influence of skip connections as the main component of the encoder-decoder to recover fine-grained details that are lost during the downsampling process. The experimental results are acquired on a study area in Finland which covers a variety minerotrophic aapa mire peatlands. The results demonstrate that multi-modal architecture consistently outperforms uni-modal architectures for peatland classification. In addition, the fusion architecture with one skip connection achieved a total accuracy of $57.44 \%$. This shows $8.51 \%$ accuracy improvement compared with the model without skip connections.

Recently, Convolutional Neural Network (CNN) has shown higher performance than other machine learning methods for land classification. In this paper, we propose a CNN fusion architecture for peatland site type classification by combining multisource and multiresolution data. The data is acquired by optical and radar satellite remote sensing, airborne laser scanning data and multi-source forest inventory GIS datasets. Based on our data, we are dealing with the high-dimensional class-imbalanced dataset for solving pixel-wise classification of peatlands. To reduce the data dimension and find an optimal subset of inputs, we first applied the sequential feature selection method. Then, we proposed a window-based pixel classification approach based on the selected inputs. This approach can extract the spatial information around each training sample in a defined window region and produce a pixel-wise classification map. Experiments are carried out for ecological classification of peatlands in Finland.

The goal of this paper is answer to this question: how much fusing LiDAR and color images can improve the performance of a convolutional neural network (CNN)-based detector? To this end, we trained state-of-the-art CNN-based detectors using different configurations of color images and their associated LiDAR data, in conjunction and independently. Moreover, we investigate the effect of sparse and dense LiDAR data on the detection accuracy. For this purpose, we estimate a dense depth image from spare LiDAR data using a recent self-supervised depth completion technique [1] that requires only sequences of color and sparse depth images, without the need for dense depth labels. Then, we compared two detectors when are trained on sparse or dense LiDAR data. The obtained results on the KITTI dataset show that fusing dense LiDAR and color images is an efficient solution for future object detectors.

The object detection task which attempts to predict bounding boxes of all interest objects in an RGB image is of paramount importance for many real-world applications and has attracted much attention within the computer vision community. However, RBG cameras cannot directly provide depth information and RGB-based object detector can not achieve an accurate performance under complex environment. To address this problem, we make two contributions in this paper. Firstly, the performances of four state-of-the art unsupervised depth estimation methods were thoroughly evaluated in the context of object detection, which can serve as a baseline for other researchers to develop even more sophisticated methods. Secondly, we investigated whether fusing depth information and RGB can improve the performance of object detection networks. The obtained results on the KITTI dataset show that RGB-depth fusion approach with MonoDepth as depth estimation method outperforms the RGB-based and depth-based detectors.

Event-based cameras are vision devices that transmit only brightness changes with low latency and ultra-low power consumption. Such characteristics make event-based cameras attractive in the field of localization and object tracking in resource-constrained systems. Since the number of generated events in such cameras is huge, the selection and filtering of the incoming events are beneficial from both increasing the accuracy of the features and reducing the computational load. In this paper, we present an algorithm to detect asynchronous corners form a stream of events in real-time on embedded systems. The algorithm is called the Three Layer Filtering-Harris or TLF-Harris algorithm. The algorithm is based on an events' filtering strategy whose purpose is 1) to increase the accuracy by deliberately eliminating some incoming events, i.e., noise and 2) to improve the real-time performance of the system, i.e., preserving a constant throughput in terms of input events per second, by discarding unnecessary events with a limited accuracy loss. An approximation of the Harris algorithm, in turn, is used to exploit its high-quality detection capability with a low-complexity implementation to enable seamless real-time performance on embedded computing platforms. The proposed algorithm is capable of selecting the best corner candidate among neighbors and achieves an average execution time savings of 59% compared with the conventional Harris score. Moreover, our approach outperforms the competing methods, such as eFAST, eHarris, and FA-Harris, in terms of real-time performance, and surpasses Arc* in terms of accuracy.

Game analytics supports game development by providing direct quantitative feedback about player experience. Player retention and monetization have become central business statistics in free-to-play game development. Total playtime and lifetime value in particular are central benchmarks, but many metrics have been used for this purpose. However, game developers often want to perform analytics in a timely manner before all users have churned from the game. This causes data censoring, which makes many metrics biased. In this article, we introduce how the mean cumulative function (MCF) can be used to measure metrics from censored data. Statistical tools based on the MCF allow game developers to determine whether a given change improves a game or whether a game is good enough for public release. The MCF is a general tool that estimates the expected value of a metric for any data set and does not rely on a model for the data. We demonstrate the advantages of this approach on a real in-development free-to-play mobile game Hipster Sheep.

Visual odometry (VO) is one of the most challenging techniques in computer vision for autonomous vehicle/vessels. In VO, the camera pose that also represents the robot pose in ego-motion is estimated analyzing the features and pixels extracted from the camera images. Different VO techniques mainly provide different trade-offs among the resources that are being considered for odometry, such as camera resolution, computation/communication capacity, power/energy consumption, and accuracy. In this paper, a hybrid technique is proposed for camera pose estimation by combining odometry based on triangulation using the long-term period of direct-based odometry and the short-term period of inverse depth mapping. Experimental results based on the EuRoC data set shows that the proposed technique significantly outperforms the traditional direct-based pose estimation method for Micro Aerial Vehicle (MAV), keeping its potential negative effect on performance negligible.

Abstract Machine learning based classification methods are widely used in geoscience applications, including mineral prospectivity mapping. Typical characteristics of the data, such as small number of positive instances, imbalanced class distributions and lack of verified negative instances make ROC analysis and cross-validation natural choices for classifier evaluation. However, recent literature has identified two sources of bias, that can affect reliability of area under ROC curve estimation via cross-validation on spatial data. The pooling procedure performed by methods such as leave-one-out can introduce a substantial negative bias to results. At the same time, spatial dependencies leading to spatial autocorrelation can result in overoptimistic results, if not corrected for. In this work, we introduce the spatial leave-pair-out cross-validation method, that corrects for both of these biases simultaneously. The methodology is used to benchmark a number of classification methods on mineral prospectivity mapping data from the Central Lapland greenstone belt. The evaluation highlights the dangers of obtaining misleading results on spatial data and demonstrates how these problems can be avoided. Further, the results show the advantages of simple linear models for this classification task.

In this paper, we propose an effective object detection framework based on proposal fusion of multiple sensors such as infrared camera, RGB cameras, radar and LiDAR. Our framework first applies the Selective Search (SS) method on RGB image data to extract possible candidate proposals which likely contain the objects of interest. Then it uses the information from other sensors in order to reduce the number of generated proposals by SS and find more dense proposals. Finally, the class of objects within the final proposals are identified by Convolutional Neural Network (CNN). Experimental results on real dataset demonstrate that our framework can precisely detect meaningful object regions using a smaller number of proposals than other object proposals methods. Further, our framework can achieve reliable object detection and classification results in maritime environments.

Maximizing product use is a central goal of many businesses, which makes retention and monetization two central analytics metrics in games. Player retention may refer to various duration variables quantifying product use: total playtime or session playtime are popular research targets, and active playtime is well-suited for subscription games. Such research often has the goal of increasing player retention or conversely decreasing player churn. Survival analysis is a framework of powerful tools well suited for retention type data. This paper contributes new methods to game analytics on how playtime can be measured using survival analysis. These methods include visualizations, metrics, and an AB-test based on the survival curve. All these methods work on censored data and enable computation of confidence intervals. This is especially important in time- and sample-limited data that occurs during game development. Throughout this paper, we illustrate the application of these methods to the development of the Hipster Sheep mobile game.

In machine learning, one often assumes the data are independent when evaluating model performance. However, this rarely holds in practice. Geographic information datasets are an example where the data points have stronger dependencies among each other the closer they are geographically. This phenomenon known as spatial autocorrelation (SAC) causes the standard cross validation (CV) methods to produce optimistically biased prediction performance estimates for spatial models, which can result in increased costs and accidents in practical applications. To overcome this problem, we propose a modified version of the CV method called spatial k-fold cross validation (SKCV), which provides a useful estimate for model prediction performance without optimistic bias due to SAC. We test SKCV with three real-world cases involving open natural data showing that the estimates produced by the ordinary CV are up to 40% more optimistic than those of SKCV. Both regression and classification cases are considered in our experiments. In addition, we will show how the SKCV method can be applied as a criterion for selecting data sampling density for new research area.

This paper proposes a digital camera based swimming analysis system for athletic use with a low budget. The recreational usage is possible during the analysis phase, and no alterations of the pool environment are needed. The system is of minimum complexity, has a real-time feedback mode, uses only underwater cameras, is flexible and can be installed in many types of public swimming pools. Possibly inaccurate camera placement poses no problem. Both commercially available and tailor made software were utilized for video signal collection and computational analysis and for providing a fast visual feedback for swimmers to improve the athletic performance. The small number of cameras with a narrow overlapping view makes the conventional stereo calibration inaccurate and a direct planar calibration method is proposed in this paper instead. The calibration method is presented and its accuracy is evaluated. The quick feedback is a key issue in improving the athletic performance. We have developed two indicators, which are easy to visualize. The first one is the swimming speed measured from the video signal by tracking a marker band at the waist of the swimmer, another one is the rudimentary swimming cycle analysis focusing to the regularity of the cycle.

A central characteristic of soil in the arctic is its load bearing capacity since that property influences forest harvester mobility, flooding dynamics and infrastructure potential. The hydraulic conductivity has the greatest dynamical influence to bearing capacity and hence is essential to measure or estimate. In addition, the arctic soil type information is needed in many cases, e.g. in roads and railways building planning. In this paper we propose a method for hydraulic conductivity estimation via linear regression on aerial gamma-ray spectroscopy and publicly available topographical data with derived elevation based features. The same data is also utilized for the arctic soil type recognition, both logistics regression and nearest neighbor classification results are reported. The classification results for logistic regression resulted in 44.5% prediction performance and 50.5% for 8-nearest neighbor classifier respectively. Linear regression results for estimating the hydraulic conductivity of the soil resulted in C-index value of 0.63. The hydraulic conductivity and soil type estimation results are promising and the proposed topographic elevation features are apparently new for remote sensing community and should also have a wider general interest.

This paper proposes a novel and simple local neural classifier for the recognition of mental tasks from on-line spontaneous EEG signals. The proposed neural classifier recognizes three mental tasks from on-line spontaneous EEG signals. Correct recognition is around 70%. This modest rate is largely compensated by two properties, namely low percentage of wrong decisions (below 5%) and rapid responses (every 1/2 s). Interestingly, the neural classifier achieves this performance with a few units, normally just one per mental task. Also, since the subject and his/her personal interface learn simultaneously from each other, subjects master it rapidly (in a few days of moderate training). Finally, analysis of learned EEG patterns confirms that for a subject to operate satisfactorily a brain interface, the latter must fit the individual features of the former.

We propose, in maximum likelihood sense, an optimal method for the affine fundamental matrix estimation in the presence of both Gaussian noise and outliers. It is based on weighting the squared residuals by the iteratively completed, residual posterior probabilities to be relevant. The proposed principle is also used for the covariance matrix estimation of the affine F-matrix where the novelty is in the fact that all data is used rather than the (erroneously) relevant classified matching points. The experiments on both synthetic and real data verify the optimality of the method in the sense of both false matches and Gaussian noise in data.

We have developed a simple and efficient wavelet-based technique for matching points in uncalibrated images. The results show that with the proposed method the probability for obtaining a false match practically vanishes when natural images are used. We also show how the uncertainty of the fundamental matrix can be interpreted from image matching viewpoint and explain how the disparity information incorporated in the fundamental matrix covariance can be utilized.

We present preliminary classification results for a real time brain-computer interface. Our approach seeks to build individual brain interfaces rather than universal ones. This means that the interface should adapt to its owner; as it will incorporate a neural classifier that learns user-specific features. Three feature sets extracted with Fourier transform, autoregressive models and wavelets were evaluated with early stopping MLP committee. The goal was to classify EEG patterns related to imagined hand movements and relax. The best results were obtained with the autoregressive special features. The results so far are not satisfactory for their intended use as basis for robust EEG classification but they give us valuable basis for future work.

Recurrent self-organizing map (RSOM) is studied in temporal sequence processing. RSOM includes a recurrent difference vector in each unit of the map, which allows storing temporal context from consecutive input vectors fed to the map. RSOM is a modification of the temporal Kohonen map (TKM). It is shown that RSOM learns a correct mapping from temporal sequences of a simple synthetic data, while TKM fails to learn this mapping. In addition, two case studies are presented, in which RSOM is applied to EEG based epileptic activity detection and to time series prediction with local models. Results suggest that RSOM can be efficiently used in temporal sequence processing.

This paper proposes a general framework for learning motion representations from low-level spatiotemporal features. The concept is based on a self-organizing map (SOM). The authors show how the SOM can be used for predicting object movements, and how additional information of the environment can be related to the inherent model of the movement to obtain generalized motion representations for objects. Traffic scenes are used to test the performance of the system.

A new method for analyzing air flows is proposed. This method is based on tracer gas technique, video recording and digital image processing. By assuming that the tracer gas cloud simulates the air movement one can recover the three-dimensional profile and the velocity field of an air flow by analyzing a picture sequence of a tracer gas cloud. All required stages and algorithms of this analyzing process are presented, including some experimental results.>