Experiments show that the proposed E-S2Feat method significantly outperforms baseline methods such as SuperEvent in pose estimation accuracy and achieves accuracy comparable to its artificial neural network counterpart while delivering an approximately 4.8-fold improvement in theoretical computational energy efficiency.
Abstract
Benefiting from high temporal resolution and dynamic range, event-based local feature methods have attracted increasing attention. However, event sparsity, noise, and limited texture still hinder robust local feature learning. Deploying such methods on resource-constrained platforms such as unmanned aerial vehicles also requires balancing accuracy and energy efficiency. To address these challenges, this paper proposes \textbf{E-S2Feat}, a spiking neural network framework for event-based local feature detection and description. The framework jointly optimizes local feature learning from the perspectives of feature representation and selection. First, a module-specific spiking activation mechanism preserves fine-grained structural cues and discriminative information under low-bit, energy-efficient inference, thereby improving overall representation fidelity. Furthermore, a semantic-guided feature modulation mechanism leverages semantic priors to refine keypoint response distributions and enhance local descriptor discriminability, thereby guiding the model to extract local features with greater geometric stability and stronger discriminative capability. Experiments on the ECD and EDS datasets show that the proposed method significantly outperforms baseline methods such as SuperEvent in pose estimation accuracy. It also achieves accuracy comparable to its artificial neural network counterpart while delivering an approximately 4.8-fold improvement in theoretical computational energy efficiency. Visual-inertial odometry experiments on the TUM-VIE dataset further verify the effectiveness and practical application potential of the proposed method in complete SLAM systems.
Event cameras generate asynchronous, high-frequency data streams offering spatially sparse information at lower latency than traditional cameras. In principle, these properties should be ideal for the design of control policies. However, reinforcement learning research in this field remains limited as existing approaches fail to fully exploit the sensor's properties. CNN-based methods negate the sensors benefits by aggregating events into sparse grids. This couples compute cost to sensor resolution and blurs the temporal information. Meanwhile, existing generative baselines rely on the availability of trajectory data to pretrain the model. We propose FLEET (Feature Learning from Events via Efficient Tokenization), a feature extractor that processes event sequences directly. Leveraging random Fourier features and cross-attention, our architecture compresses variable streams into fixed-size latent representations. This decouples inference cost of the feature extractor's backbone from the sensor's resolution, enabling end-to-end learning without auxiliary losses. We validate FLEET on a new, high-throughput benchmark. The results demonstrate that our sequence-based approach surpasses SOTA performance and exhibits superior robustness to variations in observation frequencies.
T. Gottwald, Maximilian Schier, Melanie Schaller et al.· 0 citations
Deepfakes have become increasingly realistic due to recent advances in face manipulation techniques, making reliable detection in unconstrained environments more challenging. Existing spatial-frequency deepfake detection methods often rely on fixed hand-crafted frequency transforms and simple fusion strategies, which may limit their adaptability and cross-dataset generalization. To address these limitations, we propose LGF-Net, a unified framework that jointly models spatial semantics and adaptive spectral cues for deepfake detection. The Frequency Representation Module employs learnable Gabor filters and a frequency-aware attention mechanism to capture manipulation-specific spectral patterns. Moreover, the Spatial Representation Module uses multi-rate dilated convolutions to model both subtle local artifacts and long-range structural inconsistencies, while a gated cross-modal fusion module integrates the two representations into a compact forensic descriptor. Experimental results on FF++ (HQ), Celeb-DF (V2), DPDC, and DFD show that LGF-Net achieves competitive intra-dataset and cross-dataset performance compared with several state-of-the-art deepfake detection methods.
A novel temporal iterative refinement (TIR) framework to obtain low-latency flow updates at high frequency with SNN-based feature encoders, and exhibits better cross-domain generalization, hinting toward the strong inductive biases of the network.
M. Humais, Hussain M. Sajwani, Sajid Javed et al.· IEEE Transactions on Image P...· 0 citations
GPE-YOLO is proposed, a robust detection framework built upon the YOLOv11 architecture that explicitly integrates multiscale edge priors to enhance feature resilience and validate the potential of GPE-YOLO for reliable deployment in real-world adverse weather scenarios.
Xiaojie Chen, Yi-Fei Zhou, Yi-Ming Zhou et al.· International Conference on...· 0 citations
Visual object detection is essential for environment perception in intelligent robots, automated assembly, unmanned inspection, and industrial detection systems. Although lightweight detectors reduce complexity through compact architectures, fixed convolutional units, and progressive downsampling, their limited scale responses and spatial-detail loss may degrade the localization of scale-varying and boundary-sensitive objects. To address this issue, this paper proposes DGMS-YOLO, an information-preserving lightweight detector built on a one-stage framework. The model improves feature representation through adaptive scale selection and information preservation. Specifically, the Dynamic Gated Multi-Scale Selection (DGMS) module extracts multi-scale features using depthwise convolution branches with different receptive fields and generates content-aware scale weights from the mean and standard deviation of input features. A temperature-scaled softmax and uniform scale prior are further introduced to prevent premature branch-weight concentration during multi-branch training. The Dual Pooling Downsampling (DPD) module combines max pooling, average pooling, and stride convolution to preserve salient responses, regional structures, and learnable semantic features during downsampling. In addition, high-frequency residual calibration estimates edge residuals from low-frequency features and applies lightweight channel gating to compensate for localization-related textures and boundary details. On PASCAL VOC 2007, DGMS-YOLO achieves $\text{7 6. 6 7 \%} \text{m A P}_{50}$ and $\text{5 5. 3 3 \%} \text{m A P}_{50: 95}$ with a parameter count comparable to YOLOv8s, improving it by 1.23 and 2.00 percentage points, respectively. These results demonstrate that dynamic scale selection and information preservation improve detection accuracy and localization quality under a parameter and storage budget comparable to YOLOv8s.
Xuebing Yue, Meng-Kui Hao, Yao Yao et al.· International Conference on...· 0 citations
