Search

The advantage of deep detection models primarily benefits from the feature representation ability of the backbone network, where down-sampling plays a key role in semantic integration. However, existing down-sampling approaches often ignore the global structural information of features, due to the usage of the small receptive field manner. To address this issue, this paper proposes a plug-and-play dual path down-sampling method (DPDM). It improves the support of backbone network for subsequent detection, through an extra large receptive field branch. Built on the traditional small receptive field channel, DPDM constructs an efficient large receptive field branch to obtain the structural information of features. Inspired from spatial-to-depth operation, it can achieve the effectiveness of a large receptive field under a conventional convolution kernel setting. The dual-path operation increases diversity of features but doesn’t emphasize the coordination between both types of features. Therefore, DPDM subsequently uses channel concatenation and point-wise convolution techniques to merge the features of two paths. Taking the advanced YOLO as benchmark, experimental evaluations of three models (YOLOX, YOLOv5, YOLOv6) on different datasets demonstrate the effectiveness of this method in improving detection accuracy.

Learning from a nonstationary data stream is challenging, as a data stream is generally considered to be endless, and the learning model is required to be constantly amended for adapting the shifting data distributions. When it meets multi-label data, the challenge would be further intensified. In this study, an adaptive online weighted multi-label ensemble learning algorithm called MLDME (multi-label learning with distribution matching ensemble) is proposed. It simultaneously calculates both the feature matching level and label matching level between any one reserved data block and the new received data block, further providing an adaptive decision weight assignment for ensemble classifiers based on their distribution similarities. Specifically, MLDME abandons the most commonly used but not totally correct underlying hypothesis that in a data stream, each data block always has the most approximate distribution with that emerging after it; thus, MLDME could provide a just-in-time decision for the new received data block. In addition, to avoid an infinite extension of ensemble classifiers, we use a fixed-size buffer to store them and design three different dynamic classifier updating rules. Experimental results for nine synthetic and three real-world multi-label nonstationary data streams indicate that the proposed MLDME algorithm is superior to some popular and state-of-the-art online learning paradigms and algorithms, including two specifically designed ones for classifying a nonstationary multi-label data stream. [ABSTRACT FROM AUTHOR]

With deep-learning-based object detection methods reaching industrial-level performance, underwater object detection has emerged as a significant application. However, it is often challenged by dense small instances and image blurring due to the water medium. In this paper, a Multi-Scale Feature Enhancement(MSFE) method is presented to address the challenges triggered by water bodies. In brief, MSFE attempts to achieve dual multi-scale information integration through the internal structural design of the basic C2F module in the Backbone network and the external global design of the feature pyramid network (FPN). For the internal multi-scale implementation, a LABNK module is constructed to address the vanishing or weakening phenomenon of fine-grained features during feature extraction. Specifically, it adopts a symmetrical structure to collaboratively capture two types of local receptive field information. Furthermore, to enhance the information integration ability between inter-layer features in FPN, a shallow feature branch is injected to supplement detailed features for the subsequent integration of multi-scale features. This operation is mainly supported by the fact that large-sized features from the shallow layer usually carry rich, fine-grained information. Taking the typical YOLOv8n as the benchmark model, extensive experimental comparisons on public underwater datasets (DUO and RUOD) demonstrated the effectiveness of the presented MSFE method. For example, taking the rigorous mAP (50:95) as an evaluation metric, it can achieve an accuracy improvement of about 2.8%. [ABSTRACT FROM AUTHOR]

Currently, detectors have made significant progress in inference speed and accuracy. However, these detectors require Non-Maximum Suppression (NMS) during the post-processing stage to eliminate redundant boxes, which limits the optimization of model inference speed. We first analyzed the reason for the dependence on NMS in the post-processing stage. The result showed that a score loss in a one-to-many label assignment leads to the presence of high-quality redundant boxes, making them difficult to remove. To realize end-to-end object detection and simplify the detection pipeline, we propose herein a mixed label assignment (MLA) training method, which uses one-to-many label assignment to provide rich supervision signals, alleviating the performance degradation, and we eliminate the need for NMS in the post-processing stage by using one-to-one label assignment. Additionally, a window feature propagation block (WFPB) is introduced, utilizing the inductive bias of images to enable feature sharing in local regions. Through these methods, we conducted experiments on the VOC and DUO datasets; our end-to-end detector MA-YOLOX achieved 66.0 mAP and 52.6 mAP, respectively, outperforming the YOLOX by 1.7 and 1.6. Additionally, our model performed faster than other real-time detectors without NMS. [ABSTRACT FROM AUTHOR]

The changes in appearance of faces, usually caused by pose, expression and illumination variations, increase data uncertainty in the task of face recognition. Insufficient training samples cannot provide abundant multi-view observations of a face. To address this issue, many pioneering works focus on generating virtual training images for better recognition performance. However, the issue also exists in a test set where a test image only conveys a split-second representation of a face and cannot cover more comprehensive features. In this paper, we propose a new face synthesis method for face recognition. In the proposed pipeline, we synthesize a virtual image using both the original image and its corresponding mirror one. Note that, we apply this technique both to the training and test sets. Then we use the newly generated training and test images to replace the original ones for face recognition. The aim is to increase the similarity between a test image and its corresponding intra-class training images. This proposed method is effective and computationally efficient. In order to verify this, we tested our system using multiple face recognition methods in terms of the recognition accuracy, based on either the synthesized images or original images. The methods used in the paper include statistical subspace learning algorithms and representation-based classification approaches. Experimental results obtained on FERET, ORL, GT, PIE and LFW show that the proposed approach improves the face recognition accuracy, especially on faces with left-right pose variations. [ABSTRACT FROM AUTHOR]