Complementary Trilateral Decoder for Fast and Accurate Salient Object Detection

Zhao, Zhirui; Xia, Changqun; Xie, Chenxi; Li, Jia

doi:10.1145/3474085.3475494

Cited by 80 publications

(40 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Large-scale 2D ISOD datasets such as DUTS [25] and XPIE [24] have extensively promoted the development of CNN-based ISOD methods [43]- [51]. However, 360 • panoramic SOD models are minimal due to insufficient objectlevel pixel-wise annotation [7], [8], [52].…”

Section: B 360 • Panoramic Modelsmentioning

confidence: 99%

View-Aware Salient Object Detection for $360^{\circ }$ Omnidirectional Image

Xia

et al. 2023

IEEE Trans. Multimedia

Self Cite

View full text Add to dashboard Cite

Section: B 360 • Panoramic Modelsmentioning

confidence: 99%

View-Aware Salient Object Detection for $360^{\circ }$ Omnidirectional Image

Xia

et al. 2023

IEEE Trans. Multimedia

Self Cite

View full text Add to dashboard Cite

“…The currently available deep learning-based methods for BAS [11] tend to adopt the U-shape structure proposed by UNet [34]. However, the U-shape structure suffers from two problems [20]: 1) Dilution of semantic information in the process of layer feature fusion; and 2) Weak capacity for capturing global contextual information. These problems may lead to inaccurate locations of salient objects and incomplete segmentation results.…”

Section: B Positioning Modulementioning

confidence: 99%

“…These networks first acquire a coarse saliency map, then refine it by gradually combining spatialdetail-rich features at the lower layers, and finally output the final saliency map at the top layer [12]. Zhou et al [20] created CTDNet comprising three branches with three fusion modules from coarse to fine to improve the localization of a region and the quality of its boundary. They proposed a boundary refinement module (BRM) to utilize the salient boundary information provided by the boundary path to further refine the boundary.…”

mentioning

confidence: 99%

BASNet: Burned Area Segmentation Network for Real-Time Detection of Damage Maps in Remote Sensing Images

Liu

Fan

et al. 2022

IEEE Trans. Geosci. Remote Sensing

View full text Add to dashboard Cite

Since remote sensing images of post-fire vegetation are characterized by high resolution, multiple interferences, and high similarities between the background and the target area, areas of the 2022 Beijing Winter Olympics. Fig. 1(b) shows 69 the damage map after a fire in South Korea. 70 There are many challenges for the effective retrieval of 71 burned areas from UAV images. In addition, it is important to 72 note that in the task of burned area segmentation (BAS), the 73 main purpose is to detect the whole area that has been burned 74 by fire using UAVs, which is of vital importance for post-75 disaster assessment and management. Although not belonging 76 to the burned area, new growth and unburned soil are still 77 within the scope of burned area, and thus it should be classified 78 as target burned area that is useful for further assessment 79 (The annotation of existing datasets also follows the same 80 principle). For examples, Fig. 2 shows some examples of 81 the saliency maps of different methods applied on the two 82 datasets. The first challenge is the large scale of the burned 83 area. As shown in Fig. 3(a), due to shooting time, weather 84 or geography, there are many different interferences on a 85 large area causing occlusions. Previous related studies have 86 used semantic segmentation, which in fact is a pixel-level 87 images, then longer time and more equipment are required, 146 rendering them less practical. Therefore, Tran et al. [11] used 147 cropping to process such high-resolution images with some 148 success. They cropped the images for prediction, and then 149 combined the individual results to obtain the final prediction 150 results. This resulted in two additional problems. First, the 151 burned area caused by forest fires is generally large, and the 152 segmented images may either include all burned areas or all 153 background areas in a single image. Thus, the global spatial 154 and semantic information could be lost when the deep learning 155 network performs segmentation, hence reducing the accuracy 156 of segmentation. Second, to identify large scale regions in high 157 resolution images, inputting them to the deep learning network 158 for prediction and followed by merging them require additional 159 pre-processing and post-processing. Thus, the existing BAS 160 methods cannot deal with RS images with high resolution 161 effectively. 162 In this paper, we address the challenges existing in the task 163 of segmenting post-fire burned area using RS images. The 164 major contributions of this paper are summarized as follows: 165 1) An end-to-end network BASNet based on SOD is pro-166 posed to segment the burned area using high-resolution images 167 acquired from UAV. BASNet effectively solves the problem 168 that previous methods used for BAS cannot accurately detect 169 the target in real time.170 2) The positioning and refinement modules are proposed 171 to capture the multi-level feature information by fusing rich 172 semantics with the information of spatial location and edge, 173...

show abstract

“…We compare our proposed PGNet with 11 SOTA methods, including CPD [34], SCRN [35], DASNet [43], F3Net [32], GCPA [4], ITSD [46], LDF [33], CTD [45], PFS [23], HRSOD [40], DHQSOD [30], where HRSOD and DHQ-SOD are designed for high-resolution salient object detection. All of the above methods use Resnet-50 [10] as the backbone except for HRSOD which uses VGG16 [28].…”

Section: Comparison With the State-of-the-artsmentioning

confidence: 99%

Pyramid Grafting Network for One-Stage High Resolution Saliency Detection

Xie¹,

Xia²,

Ma³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Recent salient object detection (SOD) methods based on deep neural network have achieved remarkable performance. However, most of existing SOD models designed for low-resolution input perform poorly on high-resolution images due to the contradiction between the sampling depth and the receptive field size. Aiming at resolving this contradiction, we propose a novel one-stage framework called Pyramid Grafting Network (PGNet), using transformer and CNN backbone to extract features from different resolution images independently and then graft the features from transformer branch to CNN branch. An attention-based Cross-Model Grafting Module (CMGM) is proposed to enable CNN branch to combine broken detailed information more holistically, guided by different source feature during decoding process. Moreover, we design an Attention Guided Loss (AGL) to explicitly supervise the attention matrix generated by CMGM to help the network better interact with the attention from different models. We contribute a new Ultra-High-Resolution Saliency Detection dataset UHRSD, containing 5,920 images at 4K-8K resolutions. To our knowledge, it is the largest dataset in both quantity and resolution for high-resolution SOD task, which can be used for training and testing in future research. Sufficient experiments on UHRSD and widely-used SOD datasets demonstrate that our method achieves superior performance compared to the state-of-the-art methods.

show abstract

Complementary Trilateral Decoder for Fast and Accurate Salient Object Detection

Cited by 80 publications

References 42 publications

View-Aware Salient Object Detection for $360^{\circ }$ Omnidirectional Image

View-Aware Salient Object Detection for $360^{\circ }$ Omnidirectional Image

BASNet: Burned Area Segmentation Network for Real-Time Detection of Damage Maps in Remote Sensing Images

Pyramid Grafting Network for One-Stage High Resolution Saliency Detection

Contact Info

Product

Resources

About