TVConv: Efficient Translation Variant Convolution for Layout-aware Visual Processing

Chen, Jierun; He, Tianlang; Zhuo, Weipeng; Ma, Li; Ha, Sangtae; Chan, Sixian

doi:10.1109/cvpr52688.2022.01222

Cited by 70 publications

(74 citation statements)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Chen et al [26] proposed a dynamic region-aware convolution based on the characteristics of face recognition tasks. Furthermore, in reference [27], a more comprehensive pixel-level dynamic convolution is proposed.…”

Section: Deep Learning-based Metodsmentioning

confidence: 99%

A dynamic alignment and illumination‐aware convolution for shadow removal

Wang,

Dai,

Chen

et al. 2023

IET Image Processing

View full text Add to dashboard Cite

Shadow removal is a challenging task because the variety of shadows is influenced by surface texture and lighting. This paper proposes a dynamic alignment and illumination‐aware convolution (DAIC), which consists of a Feature Alignment Module (FAM) and a Dynamic Weight Module (DWM). FAM aligns the downsampled deep features with the original features and helps to extract the optimal local information to ensure that the object texture features are not corrupted. DWM generates weights according to different lighting variations for a better shadow removal result. The shadow removal approach is based on an image decomposition algorithm using a multi‐exposure image fusion model. Here, the shadow removal network and refinement network use U‐Net framework, and the transposed convolution operations are replaced with DAIC in the decoder part of U‐Net to improve the performance of the network. The experiments are conducted on two large shadow removal datasets, ISTD+ and SRD. Compared to state‐of‐the‐art methods, this model achieves optimal performance in terms of Root Mean Square Error (RMSE) for the non‐shadow region. It also achieves performance comparable to the state‐of‐the‐art method in terms of RMSE for the shadow region and structural similarity index measurement for the entire image.

show abstract

Section: Deep Learning-based Metodsmentioning

confidence: 99%

A dynamic alignment and illumination‐aware convolution for shadow removal

Wang,

Dai,

Chen

et al. 2023

IET Image Processing

View full text Add to dashboard Cite

show abstract

“…These methods utilize depthwise convolution (DWConv) [26] and group convolution (GConv) [27] to extract spatial features. While this approach helps reduce the number of floating-point operations (FLOPs), it does not effectively address the issue of memory access reduction [28]. Previous works [28,29] have suggested minimizing redundancy in the feature maps.…”

Section: Rsconvmentioning

confidence: 99%

“…While this approach helps reduce the number of floating-point operations (FLOPs), it does not effectively address the issue of memory access reduction [28]. Previous works [28,29] have suggested minimizing redundancy in the feature maps. Inspired by this observation, we designed the segmentation-based RSConv method, illustrated in Figure 4.…”

Section: Rsconvmentioning

confidence: 99%

MAS-Net:A Novel Malaria Cell Detection Algorithm

Xiong,

2023

Preprint

View full text Add to dashboard Cite

Malaria is one of the major global health threats. Microscopy examination is the ”gold standard” for malaria diagnosis designated by the World Health Organization. However, microscopy heavily relies on doctors’ expertise, leading to prolonged diagnosis times, reduced efficiency, and the risk of misdiagnosis. This paper introduces a multi-level attention split network (MAS-Net) that addresses the challenges of information loss in detecting small targets and the discrepancy between the detection head’s receptive field and target size. Additionally, we introduce a split contextual attention structure (SPCot) that incorporates contextual information and builds a novel feature extraction network around it. SPCot avoids excessive feature map channel compression, fully utilizes contextual information, reduces information loss, and improves detection performance. Furthermore, we propose the inclusion of a multi-scale receptive field detection head (MRFH) in the shallow detection layer. This module effectively adapts to targets of varying scales, enhances the receptive fields, and improves the detection performance of malaria cells. The model achieves a significant average accuracy of 75.9% on the publicly available malaria dataset.

show abstract

“…At the same time, we found that as the number of channels increased, as shown in Figure 5, feature maps have high similarity between different channels. Based on the idea of partial convolution proposed by Chen [49], we illustrate the design of partial convolution in Figure 6. Partial convolution exploits redundancy in the feature map and systematically applies regular convolution only on a subset of input channels without affecting the remaining channels.…”

Section: Imporoved Csp Structurementioning

confidence: 99%

Badminton Hawk-Eye Challenge: A Lightweight Badminton Detection Network Based on Feature Pyramid

Jin,

Qiang,

Yan

et al. 2023

Preprint

View full text Add to dashboard Cite

The utilization of Hawk-eye technology is a reliable method to ensure the fairness and precise while playing. However, due to the high hardware cost such an application in amateur badminton matches is nearly impossible. To address this problem, this study proposes a convolutional neural network (CNN) based on the feature pyramid structure for badminton place’s detection. The model incorporates the improved MobileNetV3 structure with the CBAM mixed attention mechanism, the improved CSP structure designed to reduce redundant computation, and reach low memory access through redundant feature information. Its objective is that we aim to develop a lightweight model. Furthermore, we have incorporated the NWD loss function and frame difference to improve the accuracy of the model. Morphological processing is then employed to divide the outline of the badminton with the boundary line of the field and determine whether it is out of bounds. The model is experimentally evaluated on the badminton dataset created by us. The results indicate that the precise reaches 82.8$\%$, and the map0.5 reaches 76.8$\%$ of the compared YOLOv5 algorithm. Furthermore, the GFLOPs and parameter quantity of the proposed model are reduced to 34.9$\%$ and 50.46$\%$ of the compared YOLOv5 algorithm. Based on which we designed a challenge system for amateur badminton enthusiasts. Our lightweight Hawk-Eye challenge system has much lower hardware costs than the existing Hawk-Eye Challenge system, and is flexible in hardware deployment.

show abstract

TVConv: Efficient Translation Variant Convolution for Layout-aware Visual Processing

Cited by 70 publications

References 43 publications

A dynamic alignment and illumination‐aware convolution for shadow removal

A dynamic alignment and illumination‐aware convolution for shadow removal

MAS-Net:A Novel Malaria Cell Detection Algorithm

Badminton Hawk-Eye Challenge: A Lightweight Badminton Detection Network Based on Feature Pyramid

Contact Info

Product

Resources

About