Cross-Collaborative Fusion-Encoder Network for Robust RGB-Thermal Salient Object Detection

Multimodal salient object detection via adversarial learning with collaborative generator

“…We compare our method with 17 existing methods, including four deep learning RGBD SOD methods (i.e., S2MA , BBSnet (Fan et al, 2020b), A2dele (Piao et al, 2020) and DMRA (Piao et al, 2019)), and three traditional RGBT SOD methods (i.e., MTMR (Li et al, 2018), M3S-NIR (Tu et al, 2019) and SGDL (Tu et al, 2020)), and ten deep learning based RGBT SOD methods (i.e., ADF (Tu et al, 2022b), MIDD (Tu et al, 2021), APNet Zhou et al (2022b), ECFFNet (Zhou et al, 2022a), CSRNet (Huo et al, 2022a), CGFNet (Wang et al, 2022), MIA DPD (Liang et al, 2022), OSRNet (Huo et al, 2022b), DCNet (Tu et al, 2022a), CCFENet (Liao et al, 2022)). Different from deep learning based RGBD and RGBT SOD methods, our proposed method separates the model into three subnetworks which aim to improve Precision, Recall and Fm score of the saliency maps respectively.…”

RGB depth salient object detection via cross‐modal attention and boundary feature guidance

“…The proposed model is compared with 23 state‐of‐the‐art RGB‐D SOD models, namely BBS [28], HDFN [60], ICNet [61], VST [62], DSAM [63], DCF [64], 3DNet [65], HAIN [66], CDNet [27], UTA [67], BTS [25], TTNet [68], CMDI [69], SPNet [23], SSP [24], EENet [4], DIGR [70], CCAF [3], LDCM [71], UIFN [72], C2DF [73], CCFE [26], and Swin [2]. To be fair, all corresponding saliency maps are either provided directly by the original authors or generated directly from the trained models provided by the original authors.…”

“…In contrast to RGB cameras, depth cameras can easily acquire scenes of depth maps, which depict the geometric structure, internal consistency, and illumination invariance of salient objects. Therefore, an effective integration of RGB and depth features according to their respective advantages has become an active research focus in recent years, and a variety of RGB-D SOD [23][24][25][26][27][28] methodologies have been proposed to improve detection performance. Zhou et al [23] proposed a crossenhanced integration module, which can effectively fuse crossmodal features and learn common feature representations, moreover, the fused features are propagated to the next layer to achieve the fusion of cross-layer information.…”

“…The proposed model is compared with 23 state-of-the-art RGB-D SOD models, namely BBS [28], HDFN [60], ICNet [61], VST [62], DSAM [63], DCF [64], 3DNet [65], HAIN [66], CDNet [27], UTA [67], BTS [25], TTNet [68], CMDI [69], SPNet [23], SSP [24], EENet [4], DIGR [70], CCAF [3], LDCM [71], UIFN [72], C2DF [73], CCFE [26], and Swin [2].…”

“…In contrast to RGB cameras, depth cameras can easily acquire scenes of depth maps, which depict the geometric structure, internal consistency, and illumination invariance of salient objects. Therefore, an effective integration of RGB and depth features according to their respective advantages has become an active research focus in recent years, and a variety of RGB‐D SOD [23–28] methodologies have been proposed to improve detection performance. Zhou et al.…”