The increasing frequency of global wildfires has led to the destruction of vast forests and wetlands. Non-contact remote sensing technologies provide an effective means for accurate burned area segmentation (BAS). However, existing BAS methods often treat each image independently, focusing primarily on local pixel contexts while neglecting the broader semantic consistency of burned regions across different scenes. The lack of global context modeling limits their robustness, as burned areas typically exhibit distinctive and consistent visual characteristics such as color and texture across diverse environments. To address this limitation, we propose a Self-image and Cross-image Consistency Learning (SCCL) framework, which captures both local pixel-level relationships within a single image and global semantic dependencies across multiple images. By enforcing consistent and compact representations of burned regions within and across images, SCCL enhances segmentation robustness under varying weather and terrain conditions. Additionally, to refine boundary delineation between burned and unburned areas, we introduce a Burned Edge Injector (BEI) and an Edge-Injected Decoder (EID). We further construct two large-scale BAS benchmark datasets, BAS-AUS and BAS-EUR, for comprehensive evaluation. Experiments on these benchmarks demonstrate that our method achieves state-of-the-art performance, significantly outperforming previous approaches, with MAE reduced to 0.017 and 0.016, respectively. The new BAS benchmarks and code are available at https://github.com/VisionVerse/SCCL.

Addressing degraded weather conditions plays a vital role in practical applications. Many existing restoration approaches are limited to specific weather types, which limits their applicability to different weather scenarios. Advanced technologies, encompassing Transformer and diffusion model, have been harnessed to confront this challenge. However, these methods often heighten network complexity and prolong inference duration. To this end, we present MW-ConvNet, a U-shaped convolution-based network for multi-weather restoration. Specifically, the MW-Enc block and MW-Dec block are introduced to achieve simple yet strong feature extraction, which rely entirely on traditional 2D convolution. To improve adaptability to multiple weather conditions, a prompt generation module is designed to generate a representative weather prompt at the encoder’s terminus. Drawing inspiration from style transfer, the weather prompt is used to guide the decoder learning through a progressive restoration procedure. For future high-fidelity restoration, we introduce frequency separation through wavelet pooling blocks in encoder phase and corresponding up-sampling blocks in decoder phase. The segregated treatment of low-frequency and high-frequency features curbs the loss of textural information during network computation. It also future improves the quality and accuracy of generated weather prompt. Extensive experiments demonstrate that the proposed MW-ConvNet obtains superior performance compared to state-of-the-art methods across both weather-specific and real-world restoration tasks. Significantly, our method achieves an impressive inference speed of 0.12 seconds per$256\times 256$image, outpacing transformer-based and diffusion-based models.

Source-free domain adaptive object detection (SFOD) aims to transfer models pre-trained on the source domain to the unlabeled target domain without requiring access to the source data. Most existing SFOD methods leverage pseudo-labels for self-supervised training in the target domain. We investigate the limitations of threshold techniques to obtain high-quality pseudo-labels. In response, we design the Sequential SourceFree domain adaptive Object Detection (S-SFOD) algorithm, which enhances the quality of pseudo-labels at both the image and instance levels. At the image level, we reconstruct the training dataset, prioritizing the training of images that yield more reliable pseudo-labels to help the model acquire valuable target domain knowledge in the initial training stages. At the instance level, we introduce an adaptive local-global threshold method to balance the quality and quantity of pseudo-labels by dynamically adjusting the thresholds based on the model's learning progress. By improving the quality of pseudo-labels through these complementary techniques at both the image and instance levels, we effectively transfer knowledge from the source domain to the target domain. Extensive experiments on multiple cross-domain object detection datasets demonstrate that our proposed method outperforms current state-of-the-art SFOD algorithms. The code and model will be released.

RGB-Thermal semantic segmentation (SS) aims to combine visual light and thermal images to determine the semantic category for each pixel and create an object mask. While existing methods typically rely on well-aligned RGB-T image pairs, real-world RGB-T pairs are often unaligned, and pixel-by-pixel alignment is both challenging and time-consuming. To address this critical issue, we introduce a new unaligned RGB-T SS benchmark and propose the deformation-resilient multigranularity learning (DML) method. DML explores the spatial consistency and modal complementarity of RGB-T and mitigates the interference of warped modalities by aligning multimodal features in a coarse-to-fine multigranularity strategy. Specifically, DML constructs a deformation-aware complementary feature enhancer (DCFE), which consists of deformation-aware feature alignment (DFA) and complementary feature aggregation (CFA) modules. DFA enhances the spatial alignment of RGB-T by estimating the deformation field of warped features. Then, CFA aggregates complementary contexts of modal differences across multiple scales to produce deformation-resilient and robust RGB-T feature representations. Finally, we design the multigranularity mask refinement engine (MMFE), which combines class-agnostic saliency prediction (CSP) and class-aware edge generation (CEG) auxiliary tasks to provide useful boundary and positional cues for SS decoders. The MMFE enhances semantic alignment and interclass separability, yielding object masks with sharp boundaries. Quantitative and qualitative experiments on aligned and unaligned datasets validate the effectiveness of our proposed DML, consistently outperforming existing methods designed for aligned RGB-T data. The new unaligned RGB-T SS benchmark and code are available at https://github.com/VisionVerse/Unaligned-RGBT-Semantic-Segmentation.

Despite the tremendous success of deep neural networks (DNNs) across various fields, their susceptibility to potential backdoor attacks seriously threatens their application security, particularly in safety-critical or security-sensitive ones. Given this growing threat, there is a pressing need for research into purging backdoors from DNNs. However, prior efforts on erasing backdoor triggers not only failed to withstand increasingly powerful attacks but also resulted in reduced model performance. In this paper, we propose From Toxic to Trustworthy (FTT), an innovative approach to eliminate backdoor triggers while simultaneously enhancing model accuracy. Following the stringent and practical assumption of limited availability of clean data, we introduce a self-attention distillation (SAD) method to remove the backdoor by aligning the shallow and deep parts of the network. Furthermore, we first devise a semi-supervised learning (SSL) method that leverages ubiquitous and available poisoned data to further purify backdoors and improve accuracy. Extensive experiments on various attacks and models have shown that our FTT can reduce the attack success rate from 97% to 1% and improve the accuracy of 4% on average, demonstrating its effectiveness in mitigating backdoor attacks and improving model performance. Compared to state-of-the-art (SOTA) methods, our FTT can reduce the attack success rate by 2 times and improve the accuracy by 5%, shedding light on backdoor cleansing.

Abstract Adversarial attack for time-series classification model is widely explored and many attack methods are proposed. But there is not a method of attack based on the data itself. In this paper, we innovatively proposed a black-box sparse attack method based on data location. Our method directly attack the sensitive points in the time-series data according to statistical features extract from the dataset. At first, we have validated the transferability of sensitive points among DNNs with different structures. Secondly, we use the statistical features extract from the dataset and the sensitive rate of each point as the training set to train the predictive model. Then, predicting the sensitive rate of test set by predictive model. Finally, perturbing according to the sensitive rate. The attack is limited by constraining the L0 norm to achieve one-point attack. We conduct experiments on several datasets to validate the effectiveness of this method.

Adverse weather conditions in real-world scenarios lead to performance degradation of deep learning-based detection models. A well-known method is to use image restoration methods to enhance degraded images before object detection. However, how to build a positive correlation between these two tasks is still technically challenging. The restoration labels are also unavailable in practice. To this end, taking the hazy scene as an example, we propose a union architecture BAD-Net that connects the dehazing module and detection module in an end-to-end manner. Specifically, we design a two-branch structure with an attention fusion module for fully combining hazy and dehazing features. This reduces bad impacts on the detection module when the dehazing module performs poorly. Besides, we introduce a self-supervised haze robust loss that enables the detection module to deal with different degrees of haze. Most importantly, an interval iterative data refinement training strategy is proposed to guide the dehazing module learning with weak supervision. BAD-Net improves further detection performance through detection-friendly dehazing. Extensive experiments on RTTS and VOChaze datasets show that BAD-Net achieves higher accuracy compared to the recent state-of-the-art methods. It is a robust detection framework for bridging the gap between low-level dehazing and high-level detection.

Salient object detection (SOD) is an important task in computer vision that aims to identify visually conspicuous regions in images. RGB-Thermal SOD combines two spectra to achieve better segmentation results. However, most existing methods for RGB-T SOD use boundary maps to learn sharp boundaries, which lead to sub-optimal performance as they ignore the interactions between isolated boundary pixels and other confident pixels. To address this issue, we propose a novel position-aware relation learning network (PRLNet) for RGB-T SOD. PRLNet explores the distance and direction relationships between pixels by designing an auxiliary task and optimizing the feature structure to strengthen intra-class compactness and inter-class separation. Our method consists of two main components: A signed distance map auxiliary module (SDMAM), and a feature refinement approach with direction field (FRDF). SDMAM improves the encoder feature representation by considering the distance relationship between foreground-background pixels and boundaries, which increases the inter-class separation between foreground and background features. FRDF rectifies the features of boundary neighborhoods by exploiting the features inside salient objects. It utilizes the direction relationship of object pixels to enhance the intra-class compactness of salient features. In addition, we constitute a transformer-based decoder to decode multispectral feature representation. Experimental results on three public RGB-T SOD datasets demonstrate that our proposed method not only outperforms the state-of-the-art methods, but also can be integrated with different backbone networks in a plug-and-play manner. Ablation study and visualizations further prove the validity and interpretability of our method.

In this article, we propose a simulated crowd counting dataset CrowdX, which has a large scale, accurate labeling, parameterized realization, and high fidelity. The experimental results of using this dataset as data enhancement show that the performance of the proposed streamlined and efficient benchmark network ESA-Net can be improved by 8.4%. The other two classic heterogeneous architectures MCNN and CSRNet pre-trained on CrowdX also show significant performance improvements. Considering many influencing factors determine performance, such as background, camera angle, human density, and resolution. Although these factors are important, there is still a lack of research on how they affect crowd counting. Thanks to the CrowdX dataset with rich annotation information, we conduct a large number of data-driven comparative experiments to analyze these factors. Our research provides a reference for a deeper understanding of the crowd counting problem and puts forward some useful suggestions in the actual deployment of the algorithm.

In the field of crowd counting, the current mainstream CNNbased regression methods simply extract the density information of pedestrians without finding the position of each person. This makes the output of the network often found to contain incorrect responses, which may erroneously estimate the total number and not conducive to the interpretation of the algorithm. To this end, we propose a Bi-Branch Attention Network (BBA-NET) for crowd counting, which has three innovation points. i) A two-branch architecture is used to estimate the density information and location information separately. ii) Attention mechanism is used to facilitate feature extraction, which can reduce false responses. iii) A new density map generation method combining geometric adaptation and Voronoi split is introduced. Our method can integrate the pedestrian’s head and body information to enhance the feature expression ability of the density map. Extensive experiments performed on two public datasets show that our method achieves a lower crowd counting error compared to other state-of-the-art methods.