An improved feature pyramid network for object detection

Zhu, Linxiang; Lee, Feifei; Cai, Jiawei; Yu, Hongliu; Qiu, Chen

doi:10.1016/j.neucom.2022.02.016

Cited by 55 publications

(16 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Other researchers used stacked pyramid networks to improve the perceptual domain of human skeletal joint points and fuse the characteristics of different body parts by cascading them to improve the extension of action categories [16,17]. In the literature [18], to solve the problem of inconsistent scales of multiperson action characteristics, the authors proposed a feature Atlas preprocessing method to effectively resolve the differences between action characteristics.…”

Section: Related Workmentioning

confidence: 99%

[Retracted] Deep Learning‐Based Detection and Identification Method for Sports Health Video Dissemination

Pang

2022

Discrete Dynamics in Nature and Society

View full text Add to dashboard Cite

Sports health is gradually attracting attention, and computer vision technology is integrated into sports health to improve the quality of sports and increase the motivation of athletes. A deep learning sports health video propagation detection and recognition system is built through the mode of video propagation to provide real-time training information for sports and scientific body index parameters and exercise data for sports health programs. An athletics action estimation network (AAEN) is promoted, which initially obtains the correlation features and depth features between human skeleton key points through partial perception units. Then, all the joint point features are classified and correlated based on the affinity field range through the confidence map of the human skeletal node region. All video frames are then fused with similar joint features at the temporal level to extract motion key points in the time scale, and human posture prediction is achieved by fitting between the motion features and the dynamic database. To show the high efficiency of our method, we select three main databases for validation, and the results prove that AAEN outperforms by 13.96%, 16.90%, and 15.10% in precision, F1 score, and recall compared to the SOTA in sports health video detection and recognition. Our method also performs better overall in the same type of algorithms.

show abstract

Section: Related Workmentioning

confidence: 99%

[Retracted] Deep Learning‐Based Detection and Identification Method for Sports Health Video Dissemination

Pang

2022

Discrete Dynamics in Nature and Society

View full text Add to dashboard Cite

show abstract

“…As expected, the overall performance is promoted from 37.8 AP to 38.8 AP. The result is comparable or superior to other methods such as Libra R‐CNN [21], AugFPN [20], CARAFE [22], PANet [2] and ImFPN [61].…”

Section: Resultsmentioning

confidence: 73%

“…For the experiments on PASCAL VOC dataset, we resize the shorter dimension of the input images to 600 pixels and other hyper‐parameters are the same as in the COCO experiment. As presented in Table 3, our DCIFPN can achieve ∼ 1 AP improvement and also outperform ImFPN [61] by 2.6 AP.…”

Section: Resultsmentioning

confidence: 88%

DCIFPN: Deformable cross‐scale interaction feature pyramid network for object detection

Xiao

Jiang

et al. 2023

IET Image Processing

View full text Add to dashboard Cite

Exploiting multi-scale features is one of the most effective methods to recognize objects of different scales in object detection. Since image pyramid is time-consuming, Feature Pyramid Network (FPN) becomes the most popular component used for obtaining pyramidal features. Despite its effectiveness, there still exist some intrinsic defects. In this work, it is attributed to insufficient information flow and a Deformable Cross-scale Interaction Feature Pyramid Network (DCIFPN) is proposed, which aims to promote the information transfer process with content-aware sampling and dynamic aggregation weights. More specifically, Deformable Semantic Enhancement Module (DSEM) is designed that can construct accurate information flow with dynamic aggregation weights. In addition, Deformable Spatial Refinement Module (DSRM) is proposed to enhance high-level features with low-level location details. When DCIFPN is deployed on RetinaNet and FCOS with ResNet-50, the performance is improved by 1.6 AP and 1.1 AP, respectively, on the challenging MS COCO benchmark. Apart from one-stage detectors, DCIFPN is also applicable to two-stage methods such as Faster R-CNN and Mask R-CNN. Further experiments on Pascal VOC and CrowdHuman datasets can verify the effectiveness and generalization of the method.

show abstract

“…First, to maintain the running speed, we inherit the single‐stage detector pipeline and design an improved version based on RetinaNet [24]. Different from the common feature pyramid design, we keep the feature maps of the last three stages of the pyramid at a scaling ratio of 1/8 to capture the complete features inspired by [25, 26]. Then, to mitigate the disturbance of smoke in images, we propose a dynamic attention strategy to discover the potential features to unify scale‐awareness and spatial awareness.…”

Section: Introductionmentioning

confidence: 99%

Dynamic attention network for flame detection

Fang

Zhang

et al. 2023

Electronics Letters

View full text Add to dashboard Cite

Recently, flame detection has attracted great attention. However, existing methods have the issues of low detection rates, high false alarm rates, and lack of smoke anti-interference ability. In this letter, a novel dynamic attention-based network (DANet) is proposed for autonomous flame detection in various scenarios. To mitigate the disturbance of smoke in images, a dynamic attention strategy is proposed to discover the potential features among scale-awareness and spatial-awareness. Then, based on dynamic attention module, a decoupled detection head is presented, which can predict category, regression, and object score independently to boost the performance. A self-contained challenging flame dataset, which is multi-scene, multiscale, and multi-interference informative is constructed to evaluate the proposed model and organize the experiments. Extensive ablation and comparison studies on self-labelled dataset reveal the effectiveness of the proposed dynamic attention-based network.

show abstract

An improved feature pyramid network for object detection

Cited by 55 publications

References 16 publications

[Retracted] Deep Learning‐Based Detection and Identification Method for Sports Health Video Dissemination

[Retracted] Deep Learning‐Based Detection and Identification Method for Sports Health Video Dissemination

DCIFPN: Deformable cross‐scale interaction feature pyramid network for object detection

Dynamic attention network for flame detection

Contact Info

Product

Resources

About