Shuo Hu

Temporal action localization in long-term untrimmed videos remains a critical yet challenging task in video understanding, with existing methods often relying on anchor-based or fully-supervised frameworks that incur heavy computation and require labor-intensive frame-level annotations. This paper presents a novel weakly-supervised approach, CL-WTAL, which leverages multi-scale contrast learning and graph convolution for accurate action localization and recognition. The method comprises three key components: (i) A multi-scale sliding window mechanism (long/normal/short sequences) to segment sub-actions from complex videos, adapting to diverse action durations; (ii) A spatio-temporal graph convolution network (ST-RGCN) to extract skeletal feature vectors, integrating human motion dynamics and environmental context; (iii) A contrastive learning-based similarity evaluation framework that combines cosine similarity and Dynamic Time Warping (DTW) distance to measure feature vector relationships, enabling precise action boundary detection without extensive fine-tuning. Experiments on daily-life video datasets demonstrate that CL-WTAL effectively localizes action intervals and classifies actions with high accuracy, outperforming state-of-the-art weakly-supervised methods.

Object detection in drone imagery faces substantial challenges in achieving both high accuracy and computational efficiency for on-board deployment. To address these issues, we propose EBAD-YOLO (Efficient Bidirectional Adaptive Dense YOLO), an enhanced architecture built upon YOLOv10. Specifically, we propose a Bidirectional Adaptive Dense Connection Feature Pyramid Network (BADC-FPN) that facilitates effective multi-scale feature fusion through bidirectional cross-scale dense connections and adaptive weighting mechanisms. Building on this, we improve efficiency by optimizing C2f with the Fasterblock module. Additionally, a Localization Quality Estimation (LQE) module is incorporated into the detection head to suppress low-quality predictions, thus optimizing the precision-recall trade-off. Finally, layer-adaptive magnitude-based pruning (LAMP) is employed to further compress the model, ensuring efficient deployment. Evaluations on the VisDrone2019 dataset demonstrate that EBAD-YOLO enhances mAP@50 by 3.6 \% compared to YOLOv10s, while reducing GFLOPs by 56.5\% and the number of parameters by 66.7\%. Its robust generalization capability is further validated on the TinyPerson and LEVIR-Ship datasets.

Establishing an accurate and robust feature fusion mechanism is key to enhancing the tracking performance of single-object trackers based on a Siamese network. However, the output features of the depth-wise cross-correlation feature fusion module in fully convolutional trackers based on Siamese networks cannot establish global dependencies on the feature maps of a search area. This paper proposes a dynamic cascade feature fusion (DCFF) module by introducing a local feature guidance (LFG) module and dynamic attention modules (DAMs) after the depth-wise cross-correlation module to enhance the global dependency modeling capability during the feature fusion process. In this paper, a set of verification experiments is designed to investigate whether establishing global dependencies for the features output by the depth-wise cross-correlation operation can significantly improve the performance of fully convolutional trackers based on a Siamese network, providing experimental support for rational design of the structure of a dynamic cascade feature fusion module. Secondly, we integrate the dynamic cascade feature fusion module into the tracking framework based on a Siamese network, propose SiamDCFF, and evaluate it using public datasets. Compared with the baseline model, SiamDCFF demonstrated significant improvements.

The general paradigm of traditional Siamese networks involves using cross-correlations to fuse features from the backbone, and this paradigm is limited by the inductive bias of the convolution kernel, resulting in the lack of global information. In this paper, we propose the Siamese learning regression distribution (SiamLRD) to address the local limitations of traditional cross-correlation operations on feature fusion and weak self-connections between features within different branches. The SiamLRD uses the cross-attention mechanism to replace cross-correlations between the features of the target region of interest and the template so as to enhance flexibility. Firstly, the original transformer structure is improved to be suitable for convolutional Siamese networks. The improved transformer architecture is then used to replace cross-correlation operations, resulting in more comprehensive feature fusion between branches. Secondly, we introduce a new decoder structure into the novel fusion strategy to enhance the correlation between classification scores and regression accuracy during decoding. Multiple benchmarks are used to test the proposed SiamLRD approach, and it is verified that the proposed approach improves the baseline with 5.8% in terms of AO and 9.7% in terms of SR0.75 on the GOT-10K dataset.

Anchor-free based trackers introduce an extra branch in addition to classification and regression branches in the network to achieve comparable performance with anchor-based trackers. This extra branch is usually trained independently in the training phase and is used in combination with other branches in the inference phase. However, this can increase the inconsistency between the inference phase and the training phase, potentially degrading the tracking performance. To address this problem, we propose a new Siamese network-based object tracking framework that eliminates this inconsistency by unifying classification and additional branch tasks to achieve learning location quality estimation. Furthermore, regression tasks for bounding boxes are widely formulated based on Dirac $\delta $ distribution. Though this assumption works well for many scenarios, it restricts the prediction of regression branches. To overcome this restriction, we propose discretizing the continuous offset of the regression branch into multiple offset predictions, which enables the network to learn more flexible distributions automatically. Meanwhile, the discrete distribution prediction of regression branches is utilized to further guide the classification of the trackers. Extensive experiments on the widely accepted benchmarks demonstrate the effectiveness and efficiency of the proposed model.