Learning Selective Mutual Attention and Contrast for RGB-D Saliency Detection

SMAC

ReDWeb-S: a large-scale challenging dataset for RGB-D Salient Object Detection.

Citing our work

If you think our work is helful, please cite

@misc{liu2020learning,
      title={Learning Selective Mutual Attention and Contrast for RGB-D Saliency Detection}, 
      author={Nian Liu and Ni Zhang and Ling Shao and Junwei Han},
      year={2020},
      eprint={2010.05537},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

The Proposed RGB-D Salient Object Detection Dataset

ReDWeb-S

We construct a new large-scale challenging dataset ReDWeb-S and it has totally 3179 images with various real-world scenes and high-quality depth maps. We split the dataset into a training set with 2179 RGB-D image pairs and a testing set with the remaining 1000 image pairs.

The proposed dataset link can be found here. [baidu pan fetch code: rp8b | Google drive]

Dataset Statistics and Comparisons

We analyze the proposed ReDWeb-S datset from several statistical aspects and also conduct a comparison between ReDWeb-S and other existing RGB-D SOD datasets.

Fig.1. Top 60% scene and object category distributions of our proposed ReDWeb-S dataset.

Fig.2. Comparison of nine RGB-D SOD dataset in terms of the distributions of global contrast and interior contrast.

Fig.3. Comparsion of the average annotation maps for nine RGB-D SOD benchmark datasets.

Fig.4. Comparsion of the distribution of object size for nine RGB-D SOD benchmark datasets.

SOTA Results on our proposed dataset

We provide other SOTA RGB-D methods' results and scores on our proposed dataset. You can directly download all results [here lfa6].

Acknowledgement

We thank all annotators for helping us constructing the proposed dataset. Our proposed dataset is based on the ReDWeb dataset, which is a state-of-the-art dataset proposed for monocular image depth estimation. We also thank the authors for providing the ReDWeb dataset.

GitHub