Hyperspectral imaging (HSI) contains abundant spatial and spectral cues, making it advantageous for a wide range of applications, including Earth observation, medicine, as well as agricultural analysis. However, conventional HSI systems are often constrained by high costs, limited acquisition efficiency, and weak adaptability to dynamic scenes. To mitigate these limitations, we propose a progressive multi-feature fusion model, termed PMFF-SRNet, for RGB-to-hyperspectral reconstruction. The proposed model progressively recovers spectral information through multiple stages, which helps reduce spectral redundancy while improving reconstruction efficiency. Furthermore, a Local–Global Spectral Attention (LGSA) module is employed to model spectral features at different granularities, where grouped self-attention focuses on local band interactions while spectral order information contributes to long-range dependency modeling. In addition, a Discrete Wavelet Attention (DWA) module is incorporated into the skip connections to enhance texture and edge restoration by exploiting the multi-scale characteristics of wavelet transforms. Results obtained on the NTIRE benchmarks indicate that PMFF-SRNet achieves competitive reconstruction performance across multiple evaluation metrics, while maintaining a lightweight and computationally efficient architecture. These findings demonstrate the strong potential of PMFF-SRNet for practical hyperspectral reconstruction tasks.

High-value assets on the battlefield typically require adequate camouflage to evade detection and annihilation by enemy scouts. Consequently, artificial camouflage technology is extensively acknowledged and utilized as a crucial defensive tactic in the military sphere. The quality of camouflage performance was assessed by military observers through the human visual system (HVS). This method involved locating the camouflaged objects and rating the camouflaged degree against the background. Current camouflage assessment methods typically involved the manual extraction and aggregation of objective features throughout an image. These approaches fall short in constructing a correlation mapping between objective features and subjective perceptions of camouflaged objects, culminating in imprecise assessments and discrepancies. To address these issues, this paper presents the first three-stage full-reference learning framework for locating camouflaged objects, extracting camouflage features, and assessing camouflage quality. Given the lack of datasets specifically designed for evaluating camouflage quality, we have contributed a datasets focused on human-camouflaged targets. The experimental results show that the three-stage framework is remarkably accurate in assessing the camouflage quality, leading to an explainable network. The camouflaged people quality assessment(CPQA) dataset is available at http://github.com/samsunq/CPQA_Datasets.git . • A three-stage learning model is proposed to measure the quality of camouflaged targets, capable of simultaneous localization, feature extraction, that is derived from the human visual system (HVS). • Collected and created the first camouflaged people evaluation assessment (CPQA) dataset. • The proposed method achieves an evaluation outcome in the CPQA dataset, which is nearly linearly correlated with the Mean Opinion Scores (MOS) and surpasses current alternative evaluation methods.

Camouflaged Objects Detection (COD) aims to identify objects seamlessly blending into their surrounding environments. Existing COD methods treat COD as a binary segmentation problem based on Salient Object Detection techniques, which separate objects from the background. While these methods have been widely applied, their inability to identify object categories limits the scope of applications. Moreover, multi-target selection and localization rely heavily on expert-driven post-processing, resulting in poor interactivity. To address these limitations, we reformulate COD as a Referring Image Segmentation (RIS) challenge, enabling precise localization and segmentation of language-specified objects through natural language instructions. Accordingly, this paper proposes a novel RIS framework named MSIENet for the COD task, which integrates a language encoder, an image encoder, and a multi-modal fusion module. This framework bridges the modality gap between visual and linguistic features through a cross-attention-based fusion and alignment module. MSIENet also contains two key components: multi-scale edge enhancement and texture enhancement modules, which effectively aggregate and refine texture details and boundary information, facilitating the generation of high-quality segmentation masks. We also collect a Language-image camouflaged dataset Ref-ACOD, establishing a rigorous evaluation benchmark for COD tasks based on RIS approaches. Experiments demonstrate that the MSIENet surpasses SOTA RIS methods on COD tasks, with MIoUs and OIoUs on LAVT increasing by 8 % and 14.5 %. All datasets are available at http://github.com/samsunq/Ref-ACOD.git

Haze and fog, as severe weather conditions, have absorbing and scattering effects on the optical images, severely affecting image quality. Polarization-based dehazing algorithms can estimate the original radiance distribution of the scene through the polarization of skylight and transmitted light. However, current traditional methods lack consideration for the polarization of transmitted light, and the datasets required for deep learning-based methods are difficult to obtain. This paper proposes a polarized haze image synthesis method that can generate scene intensity and polarization after passing through different distances and concentrations from existing DoFP images, Specially, we equate the attenuation of the scattering medium to a superposition of a series of Mueller matrices, and in combination with the atmospheric attenuation model, which thoroughly integrates both intensity characteristics and polarization properties. We establish a comprehensive polarization dataset for image dehazing, including 300 sets of simulated data, 40 sets real world data from artificial scenes with haze-free ground truth and 40 sets real world data from urban scenes. The network model trained on our simulated dataset demonstrates the effectiveness of the simulation method in testing experiments.

Vision-based localization is a cost-effective and thus attractive solution for many intelligent mobile platforms. However, its accuracy and especially robustness still suffer from low illumination conditions, illumination changes, and aggressive motion. Event-based cameras are bio-inspired visual sensors that perform well in HDR conditions and have high temporal resolution, and thus provide an interesting alternative in such challenging scenarios. While purely event-based solutions currently do not yet produce satisfying mapping results, the present work demonstrates the feasibility of purely event-based tracking if an alternative sensor is permitted for mapping. The method relies on geometric 3D-2D registration of semi-dense maps and events, and achieves highly reliable and accurate cross-modal tracking results. Practically relevant scenarios are given by depth camera-supported tracking or map-based localization with a semi-dense map prior created by a regular image-based visual SLAM or structure-from-motion system. Conventional edge-based 3D-2D alignment is extended by a novel polarity-aware registration that makes use of signed time-surface maps (STSM) obtained from event streams. We furthermore introduce a novel culling strategy for occluded points. Both modifications increase the speed of the tracker and its robustness against occlusions or large view-point variations. The approach is validated on many real datasets covering the above-mentioned challenging conditions, and compared against similar solutions realised with regular cameras.

We present a novel real-time visual odometry framework for a stereo setup of a depth and high-resolution event camera. Our framework balances accuracy and robustness against computational efficiency towards strong performance in challenging scenarios. We extend conventional edge-based semi-dense visual odometry towards time-surface maps obtained from event streams. Semi-dense depth maps are generated by warping the corresponding depth values of the extrinsically calibrated depth camera. The tracking module updates the camera pose through efficient, geometric semi-dense 3D-2D edge alignment. Our approach is validated on both public and self-collected datasets captured under various conditions. We show that the proposed method performs comparable to state-of-the-art RGB-D camera-based alternatives in regular conditions, and eventually outperforms in challenging conditions such as high dynamics or low illumination.

Event-based visual perception is becoming increasingly popular owing to interesting sensor characteristics enabling the handling of difficult conditions such as highly dynamic motion or challenging illumination. The mostly complementary nature of event cameras however still means that best results are achieved if the sensor is paired with a regular frame-based sensor. The present work aims at answering a simple question: Assuming that both cameras do not share a common optical center, is it possible to exploit the hybrid stereo setup's baseline to perform accurate stereo depth estimation? We present a learning based solution to this problem leveraging modern spatio-temporal input representations as well as a novel hybrid pyramid attention module. Results on real data demonstrate competitive performance against pure frame-based stereo alternatives as well as the ability to maintain the advantageous properties of event-based sensors.