Parameter-Efficient Fine-Tuning (PEFT) is a fundamental research problem in computer vision, which aims to tune a few parameters for efficient storage and adaptation of pre-trained vision models. Recently, sensitivity-aware parameter efficient fine-tuning method (SPT) addresses this problem by identifying sensitive parameters and then leveraging its sparse characteristic to combine unstructured and structured tuning for PEFT. However, existing methods only focus on the sparse characteristic of sensitive parameters but overlook its distribution characteristic, which results in additional storage burden and limited performance improvement. In this paper, we find that the distribution of sensitive parameters is not chaotic, but concentrates on a small number of rows or columns in each parameter matrix. Inspired by this fact, we propose a Compact Dynamic-Rank Adaptation-based tuning method for Sensitivity-aware Parameter efficient fine-Tuning, called CDRA-SPT. Specifically, we first identify the sensitive parameters that require tuning for each down-stream task. Then, we reorganize the sensitive parameters by following its row and column into a compact sub-parameter matrix. Finally, a dynamic-rank adaptation is designed and applied at sub-parameter matrix level for PEFT. Its advantage is that the dynamic-rank characteristic of sub-parameter matrix can be fully exploited for PEFT. Extensive experiments show that our method achieves superior performance over previous state-of-the-art methods.

Precipitation nowcasting involves using current radar observation sequences to predict future radar sequences and determine future precipitation distribution, which is crucial for disaster warning, traffic planning, and agricultural production. Despite numerous advancements, challenges persist in accurately predicting both the location and intensity of precipitation, as these factors are often interdependent, with complex atmospheric dynamics and moisture distribution causing position and intensity changes to be intricately coupled. Inspired by the fact that in the frequency domain, phase variations are shown to correspond to changes in the position of precipitation, while amplitude variations are linked to intensity changes, we propose an amplitude-phase disentanglement model called AlphaPre, which separately learn the position and intensity changes of precipitation. AlphaPre comprises three key components: a phase network, an amplitude network, and an AlphaMixer. The phase network captures positional changes by learning phase variations, and the amplitude network models intensity changes by alternating between the frequency and spatial domains. The AlphaMixer then integrates these components to produce a refined precipitation forecast. Extensive experiments on four datasets demonstrate the effectiveness and superiority of our method over state-of-the-art approaches. Our code is publicly available at https://github.com/linkenghong/AlphaPre.

In recent years, heatmap regression models have attracted much attention due to their superior performance in locating facial landmarks. These models overcome the lack of spatial and contextual information of coordinate regression models. However, there are still two major problems with these models; (1) they are computationally expensive, (2) they have the quantization error. To address two issues, we present a lightweight facial landmark detector named LFLD that unifies ShuffleNetV2 with the principles of the high-resolution structure for facial landmark detection. Our LFLD has high accuracy and significantly reduced computation costs compared to other heatmap regression models. In addition, the LFLD is equipped with two detection heads that generate the primary heatmaps and the auxiliary heatmaps. The primary heatmaps represent the integer part of the predicted facial landmarks coordinates, and the auxiliary heatmaps indicate the fractional part. Our LFLD uses these two heatmaps to jointly represent the predicted facial landmarks coordinates. The extensive experimental results show that LFLD can achieve satisfactory results in the case of low parameters and FLOPs. Particularly for the NME, our LFLD reaches 4.31% on WFLW and 3.54% on 300W. In heatmap regression, when compared to HRNet, our LFLD has lower NME than HRNet and has 45% fewer parameters and 66% fewer FLOPs. Finally, our model has fewer parameters than large networks, and the detection accuracy is comparable to large networks.

Information at different resolutions plays distinct roles in computer vision tasks. Although the research on the utilization of different resolution information in semantic segmentation has made some achievements, the research on the utilization of different resolution information in real-time semantic segmentation needs to be improved. To address this issue, we propose Deep Multi-Resolution Network (DMRNet), a lightweight model using different resolution information for real-time semantic segmentation. This model consists of several branches with different resolutions, and information is fused between neighbouring branches after convolution operations. At the end of the lowest resolution branch, we designed an enhanced semantic information module, Amplify Aggregate Pyramid Pooling Module (AAPPM), to balance the extraction of semantic information with the speed of inference. In addition, at the end of all branches, we propose a multi-resolution fusion module (MRFM) to guide the information fusion in different branches, which helps to improve the problem of spatial details being covered by semantic information. On CityScapes and Camvid, the most widely-used datasets in the field of semantic segmentation, our method strikes a balance between network accuracy and inference speed. On a single 2080Ti GPU, DMRNet achieves 77.6 % and 74.7 % accuracy at inference speeds of 68.7 FPS and 91.6 FPS, respectively.

We present an adaptive multi-feature fusion visual object tracking algorithm based on Siamese neural network with cross-attention mechanism, SiamAtten for short, which can effectively deal with large appearance changes, complex back-ground and interference. The proposed network consists of two parts. One is the full convolution feature extraction network based on cross-attention mechanism and the other is the region proposal generation network based on adaptive multi-feature fusion. The cross-attention mechanism is used in the first part to improve the response ability of feature extraction of the target. The adaptive feature fusion method is used in the second part to infer the target location in a step-by-step process and to get the robust region proposal by regression. Meanwhile, network parameters are reduced by using the depth-wise separable convolution, and the cross-attention mechanism is proposed in this paper can effectively enhance the target identification ability and elevate the robustness. Extensive experiments are carried out on three benchmark datasets and much advanced tracking results are obtained.

Although elevating the accuracy and efficiency of facial biometric recognition system, it suffers from presentation attacks (PAs) because of its weakness. Currently, popular state- of-the-art face anti-spoofing (FAS) methods using multi-modal learning strategy. Similarly, we propose multi-modal FAS using multi-fusion network (MFN) and global depth-wise convolution (GDConv), FaceBagNetPlus for short. The MFN means that we use the Convolutional Block Attention Module (CBAM) to replace the Squeeze-and-Excitation Network (SE-NET) in the feature extraction part and propose channel spatial cross fusion (CSCF) to cross-fuse modal feature with the pairwise cross approach. Meanwhile, we use the GDConv to replace the global average pooling (GAP) to raise the performance. Then, we use the patch-based strategy to obtain fully feature, the random model feature erasing (RMFE) strategy to avoid over-fitting and multi-stream fusion module to enhance discrimination ability. Next, we perform experiments on the CASIA-SURF dataset, then demonstrate the effectiveness of the MFN and the GDConv. Among all results, we gain the best result of 113 (FP), 4 (FN), 0.2807% (APCER), 0.0229% (NPCER), 0.1518% (ACER), 100.000% (TPR@FPR=10e-2), 100.000% (TPR@FPR=10e-3) and 99.9026% (TPR@FPR=10e-4) on the test set. We also execute experiments on the CASIA-SURF CeFA dataset and receive the best result of 0.0000% (ACER) on the validation set. Finally, two results are superior to the state-of-the-art methods.