The tokenizer, as one of the fundamental components of large models, has long been overlooked or even misunderstood in visual tasks. One key factor of the great comprehension power of the large language model is that natural language tokenizers utilize meaningful words or subwords as the basic elements of language. In contrast, mainstream visual tokenizers, represented by patch-based methods such as Patch Embed, rely on meaningless rectangular patches as basic elements of vision, which cannot serve as effectively as words or subwords in language. Starting from the essence of the tokenizer, we defined semantically independent regions (SIRs) for vision. We designed a simple HOmogeneous visual tOKenizer: HOOK. HOOK mainly consists of two modules: the Object Perception Module (OPM) and the Object Vectorization Module (OVM). To achieve homogeneity, the OPM splits the image into 4*4 pixel seeds and then utilizes the attention mechanism to perceive SIRs. The OVM employs cross-attention to merge seeds within the same SIR. To achieve adaptability, the OVM defines a variable number of learnable vectors as cross-attention queries, allowing for the adjustment of token quantity. We conducted experiments on the NWPU-RESISC45, WHU-RS19 classification dataset, and GID5 segmentation dataset for sparse and dense tasks. The results demonstrate that the visual tokens obtained by HOOK correspond to individual objects, which demonstrates homogeneity. HOOK outperformed Patch Embed by 6\% and 10\% in the two tasks and achieved state-of-the-art performance compared to the baselines used for comparison. Compared to Patch Embed, which requires more than one hundred tokens for one image, HOOK requires only 6 and 8 tokens for sparse and dense tasks, respectively, resulting in efficiency improvements of 1.5 to 2.8 times. The code is available at https://github.com/GeoX-Lab/Hook.
Accurate traffic forecasting is a fundamental problem in intelligent transportation systems and learning long-range traffic representations with key information through spatiotemporal graph neural networks (STGNNs) is a basic assumption of current traffic flow prediction models. However, due to structural limitations, existing STGNNs can only utilize short-range traffic flow data; therefore, the models cannot adequately learn the complex trends and periodic features in traffic flow. Besides, it is challenging to extract the key temporal information from the long historical traffic series and obtain a compact representation. To solve the above problems, we propose a novel LSTTN (Long-Short Term Transformer-based Network) framework comprehensively considering the long- and short-term features in historical traffic flow. First, we employ a masked subseries Transformer to infer the content of masked subseries from a small portion of unmasked subseries and their temporal context in a pretraining manner, forcing the model to efficiently learn compressed and contextual subseries temporal representations from long historical series. Then, based on the learned representations, long-term trend is extracted by using stacked 1D dilated convolution layers, and periodic features are extracted by dynamic graph convolution layers. For the difficulties in making time-step level prediction, LSTTN adopts a short-term trend extractor to learn fine-grained short-term temporal features. Finally, LSTTN fuses the long-term trend, periodic features and short-term features to obtain the prediction results. Experiments on four real-world datasets show that in 60-minute-ahead long-term forecasting, the LSTTN model achieves a minimum improvement of 5.63\% and a maximum improvement of 16.78\% over baseline models. The source code is available at https://github.com/GeoX-Lab/LSTTN.
Deep neural network-based Synthetic Aperture Radar (SAR) target recognition models are susceptible to adversarial examples. Current adversarial example generation methods for SAR imagery primarily operate in the 2D digital domain, known as image adversarial examples. Recent work, while considering SAR imaging scatter mechanisms, fails to account for the actual imaging process, rendering attacks in the three-dimensional physical domain infeasible, termed pseudo physics adversarial examples. To address these challenges, this paper proposes SAR-AE-SFP-Attack, a method to generate real physics adversarial examples by altering the scattering feature parameters of target objects. Specifically, we iteratively optimize the coherent energy accumulation of the target echo by perturbing the reflection coefficient and scattering coefficient in the scattering feature parameters of the three-dimensional target object, and obtain the adversarial example after echo signal processing and imaging processing in the RaySAR simulator. Experimental results show that compared to digital adversarial attack methods, SAR-AE-SFP Attack significantly improves attack efficiency on CNN-based models (over 30\%) and Transformer-based models (over 13\%), demonstrating significant transferability of attack effects across different models and perspectives.
In the later training stages, further improvement of the models ability to determine changes relies on how well the change detection (CD) model learns hard cases; however, there are two additional challenges to learning hard case samples: (1) change labels are limited and tend to pointer only to foreground targets, yet hard case samples are prevalent in the background, which leads to optimizing the loss function focusing on the foreground targets and ignoring the background hard cases, which we call imbalance. (2) Complex situations, such as light shadows, target occlusion, and seasonal changes, induce hard case samples, and in the absence of both supervisory and scene information, it is difficult for the model to learn hard case samples directly to accurately obtain the feature representations of the change information, which we call missingness. We propose a Siamese foreground association-driven hard case sample optimization network (HSONet). To deal with this imbalance, we propose an equilibrium optimization loss function to regulate the optimization focus of the foreground and background, determine the hard case samples through the distribution of the loss values, and introduce dynamic weights in the loss term to gradually shift the optimization focus of the loss from the foreground to the background hard cases as the training progresses. To address this missingness, we understand hard case samples with the help of the scene context, propose the scene-foreground association module, use potential remote sensing spatial scene information to model the association between the target of interest in the foreground and the related context to obtain scene embedding, and apply this information to the feature reinforcement of hard cases. Experiments on four public datasets show that HSONet outperforms current state-of-the-art CD methods, particularly in detecting hard case samples.
For a long time, due to the high heterogeneity in structure and semantics among various spatiotemporal modal data, the joint interpretation of multimodal spatiotemporal data has been an extremely challenging problem. The primary challenge resides in striking a trade-off between the cohesion and autonomy of diverse modalities, and this trade-off exhibits a progressively nonlinear nature as the number of modalities expands. We introduce the Language as Reference Framework (LaRF), a fundamental principle for constructing a multimodal unified model, aiming to strike a trade-off between the cohesion and autonomy among different modalities. We propose a multimodal spatiotemporal general artificial intelligence model, called AllSpark. Our model integrates thirteen different modalities into a unified framework, including 1D (text, code), 2D (RGB, infrared, SAR, multispectral, hyperspectral, tables, graphs, trajectory, oblique photography), and 3D (point clouds, videos) modalities. To achieve modal cohesion, AllSpark uniformly maps diverse modal features to the language modality. In addition, we design modality-specific prompts to guide multi-modal large language models in accurately perceiving multimodal data. To maintain modality autonomy, AllSpark introduces modality-specific encoders to extract the tokens of various spatiotemporal modalities. And modal bridge is employed to achieve dimensional projection from each modality to the language modality. Finally, observing a gap between the model's interpretation and downstream tasks, we designed task heads to enhance the model's generalization capability on specific downstream tasks. Experiments indicate that AllSpark achieves competitive accuracy in modalities such as RGB and trajectory compared to state-of-the-art models.
We apply the U-Net model for compressive light field synthesis. Compared to methods based on stacked CNN and iterative algorithms, this method offers better image quality, uniformity and less computation.
Ranker and retriever are two important components in dense passage retrieval. The retriever typically adopts a dual-encoder model, where queries and documents are separately input into two pre-trained models, and the vectors generated by the models are used for similarity calculation. The ranker often uses a cross-encoder model, where the concatenated query-document pairs are input into a pre-trained model to obtain word similarities. However, the dual-encoder model lacks interaction between queries and documents due to its independent encoding, while the cross-encoder model requires substantial computational cost for attention calculation, making it difficult to obtain real-time retrieval results. In this paper, we propose a dense retrieval model called MD2PR based on multi-level distillation. In this model, we distill the knowledge learned from the cross-encoder to the dual-encoder at both the sentence level and word level. Sentence-level distillation enhances the dual-encoder on capturing the themes and emotions of sentences. Word-level distillation improves the dual-encoder in analysis of word semantics and relationships. As a result, the dual-encoder can be used independently for subsequent encoding and retrieval, avoiding the significant computational cost associated with the participation of the cross-encoder. Furthermore, we propose a simple dynamic filtering method, which updates the threshold during multiple training iterations to ensure the effective identification of false negatives and thus obtains a more comprehensive semantic representation space. The experimental results over two standard datasets show our MD2PR outperforms 11 baseline models in terms of MRR and Recall metrics.
Local Attention-guided Message Passing Mechanism (LAMP) adopted in Graph Attention Networks (GATs) is designed to adaptively learn the importance of neighboring nodes for better local aggregation on the graph, which can bring the representations of similar neighbors closer effectively, thus showing stronger discrimination ability. However, existing GATs suffer from a significant discrimination ability decline in heterophilic graphs because the high proportion of dissimilar neighbors can weaken the self-attention of the central node, jointly resulting in the deviation of the central node from similar nodes in the representation space. This kind of effect generated by neighboring nodes is called the Distraction Effect (DE) in this paper. To estimate and weaken the DE of neighboring nodes, we propose a Causally graph Attention network for Trimming heterophilic graph (CAT). To estimate the DE, since the DE are generated through two paths (grab the attention assigned to neighbors and reduce the self-attention of the central node), we use Total Effect to model DE, which is a kind of causal estimand and can be estimated from intervened data; To weaken the DE, we identify the neighbors with the highest DE (we call them Distraction Neighbors) and remove them. We adopt three representative GATs as the base model within the proposed CAT framework and conduct experiments on seven heterophilic datasets in three different sizes. Comparative experiments show that CAT can improve the node classification accuracy of all base GAT models. Ablation experiments and visualization further validate the enhancement of discrimination ability brought by CAT. The source code is available at https://github.com/GeoX-Lab/CAT.
Recently, deep learning-based tooth segmentation methods have been limited by the expensive and time-consuming processes of data collection and labeling. Achieving high-precision segmentation with limited datasets is critical. A viable solution to this entails fine-tuning pre-trained multiview-based models, thereby enhancing performance with limited data. However, relying solely on two-dimensional (2D) images for three-dimensional (3D) tooth segmentation can produce suboptimal outcomes because of occlusion and deformation, i.e., incomplete and distorted shape perception. To improve this fine-tuning-based solution, this paper advocates 2D-3D joint perception. The fundamental challenge in employing 2D-3D joint perception with limited data is that the 3D-related inputs and modules must follow a lightweight policy instead of using huge 3D data and parameter-rich modules that require extensive training data. Following this lightweight policy, this paper selects skeletons as the 3D inputs and introduces MSFormer, a novel method for tooth segmentation. MSFormer incorporates two lightweight modules into existing multiview-based models: a 3D-skeleton perception module to extract 3D perception from skeletons and a skeleton-image contrastive learning module to obtain the 2D-3D joint perception by fusing both multiview and skeleton perceptions. The experimental results reveal that MSFormer paired with large pre-trained multiview models achieves state-of-the-art performance, requiring only 100 training meshes. Furthermore, the segmentation accuracy is improved by 2.4%-5.5% with the increasing volume of training data.
Data-driven deep learning methods have shown great potential in cropland mapping. However, due to multiple factors such as attributes of cropland (topography, climate, crop type) and imaging conditions (viewing angle, illumination, scale), croplands under different scenes demonstrate a great domain gap. This makes it difficult for models trained in the specific scenes to directly generalize to other scenes. A common way to handle this problem is through the "Pretrain+Fine-tuning" paradigm. Unfortunately, considering the variety of features of cropland that are affected by multiple factors, it is hardly to handle the complex domain gap between pre-trained data and target data using only sparse fine-tuned samples as general constraints. Moreover, as the number of model parameters grows, fine-tuning is no longer an easy and low-cost task. With the emergence of prompt learning via visual foundation models, the "Pretrain+Prompting" paradigm redesigns the optimization target by introducing individual prompts for each single sample. This simplifies the domain adaption from generic to specific scenes during model reasoning processes. Therefore, we introduce the "Pretrain+Prompting" paradigm to interpreting cropland scenes and design the auto-prompting (APT) method based on freely available global land cover product. It can achieve a fine-grained adaptation process from generic scenes to specialized cropland scenes without introducing additional label costs. To our best knowledge, this work pioneers the exploration of the domain adaption problems for cropland mapping under prompt learning perspectives. Our experiments using two sub-meter cropland datasets from southern and northern China demonstrated that the proposed method via visual foundation models outperforms traditional supervised learning and fine-tuning approaches in the field of remote sensing.