SmallTrack: Wavelet Pooling and Graph Enhanced Classification for UAV Small Object Tracking

Xue, Yuanliang; Jin, Guodong; Shen, Tao; Tan, Lining; Wang, Nian; Gao, Jing; Lian-feng, Wang

doi:10.1109/tgrs.2023.3305728

Cited by 11 publications

(3 citation statements)

References 69 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We compare the proposed SGSiamAttn with other 24 state-of-the-art trackers, which can be divided into two categories: CNN-based trackers and transformer-based trackers. The CNN-based trackers include SiamRPN [15], SiamRPN++ [18], SiamMask [54], SiamDW [19], ATOM [55], SiamBAN [6], SiamCAR [7], SiamFC++ [24], SiamKPN [25], Ocean [21], CLNet [56], SiamRN [57], SiamGAT [22], SiamRBO [58], EnSiamMask [2], CCST [59], SiamMSS [33], and SmallTrack [23]. TransT [26], TCTrack [27], DeconNet [28], DropTrack [29], GRM [30], and ARTrack [31] make up the transformer-based trackers.…”

Section: E Comparison Experimentsmentioning

confidence: 99%

“…In addition, there are several works investigating the discriminative power of the tracker [20], and the misalignment between classification and regression [21], and alternatives to cross-correlation operations [22]. Xue et al [23] embed wavelet transform and graph neural networks into the tracking framework for emphasizing small targets. Developed trackers generate candidate boxes in an anchorbased manner, which requires the priori knowledge of the object and involves some necessary hyperparameters, such as the size and aspect ratio of the anchor.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Saliency Guided Siamese Attention Network for Infrared Ship Target Tracking

Li,

Zhang,

Liu

et al. 2024

IEEE Trans. Intell. Veh.

View full text Add to dashboard Cite

Due to the lack of discriminative features in infrared images, most of existing trackers cannot separate a target from its background. There are some studies on generating discriminative features where feature fusion and attention are applied to enhance targets. However, the saliency information and information interaction which assist in locating the targets is ignored. To improve the accuracy of infrared ship target tracking, we propose a saliency guided Siamese attention network (SGSiamAttn). The main contribution is to design a saliency prediction network that obtains the saliency map of a search region and followed by a saliency enhancement network to highlight the target. With the saliency information, our network is able to perceive the entire target, which improves the discriminative ability and the tracking accuracy. Meanwhile, a local-to-global correlation module is applied before the saliency prediction network, aiming to refine the correlation map while suppressing non-target interferences. We also impose a shared cross-correlation module on the region proposal network. By sharing the correlation map in the classification and regression branches, it enhances information interaction between the two tasks and reduces the computational cost. As there are limited number of infrared ship tracking datasets publicly available, we construct a new infrared ship dataset (ISD) which includes 16 different types of ships and 7,872 video frames with manual annotations. The experimental results on ISD and other three public datasets, namely VOT-TIR2015, PTB-TIR, and LSOTB-TIR, demonstrate that our tracker achieves superior performance in terms of accuracy, expected average overlap, success, and precision.

show abstract

Section: E Comparison Experimentsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Saliency Guided Siamese Attention Network for Infrared Ship Target Tracking

Li,

Zhang,

Liu

et al. 2024

IEEE Trans. Intell. Veh.

View full text Add to dashboard Cite

show abstract

“…This feature enhances the perceptual functions of UAV, particularly in a scenarios with limited human-drone interaction, providing fundamental technical support for autonomous detection and flight. To improve the feature quality of small targets for drones, SmallTrack 4 uses a small wave pool layer to avoid losing small target information during downsampling and highlights the classification response of small objects through a graph enhancement module, which improves detection accuracy but is difficult to meet real-time requirements. MobileTrack 5 proposes a lightweight convolutional network and enhances the feature information of targets by designing an efficient target perception module for local cross-channel information exchange, achieving a balance between speed and accuracy.…”

Section: Introductionmentioning

confidence: 99%

A new lightweight network for efficient UAV object detection

Hua,

Chen,

Chen

2024

Sci Rep

View full text Add to dashboard Cite

Optimizing the structure of deep neural networks is essential in many applications. Especially in the object detection tasks of Unmanned Aerial Vehicles. Due to the constraints of the onboard platform, a more efficient network is required to meet practical demands. Nevertheless, existing lightweight detection networks exhibit excessive redundant computations and may yield in a certain level of accuracy loss. To address these issues, this paper proposes a new lightweight network structure named Cross-Stage Partially Deformable Network (CSPDNet). The initial proposal consists of a Deformable Separable Convolution Block (DSCBlock), separating feature channels, greatly reducing the computational load of convolution, and applying adaptive sampling to the separated feature map. Subsequently, to establish information interaction between feature layers, a channel weighting module is proposed. This module calculates weights for the separated feature map, facilitating information exchange across channels and resolutions. Moreover, it compensates for the effect of point-wise (1 $$\times$$ × 1) convolutions, filtering out more important feature information. Furthermore, a new CSPDBlock is designed, primarily composed of DSCBlock, establishing multidimensional feature correlations for each separated feature layer. This approach improves the ability to capture critical feature information and reconstruct gradient paths, thereby preserving detection accuracy. The proposed technology achieves a balance between model parameter size and detection accuracy. Furthermore, experimental results on object detection datasets demonstrate that our designed network, using fewer parameters, achieves competitive detection performance results compared to existing lightweight networks YOLOv5n, YOLOv6n, YOLOv8n, NanoDet and PP-PicoDet. The optimization effect of the designed CSPDBlock, using the VisDrone dataset, is validated when incorporated into advanced detection algorithms YOLOv5m, PPYOLOEm, YOLOv7, RTMDetm and YOLOv8m. In more detail, by incorporating the designed modules it was achieved that the parameters were reduced by 10–20% while almost maintaining detection accuracy.

show abstract