/SSLSOD

(AAAI 2022) Self-Supervised Pretraining for RGB-D Salient Object Detection

Primary LanguagePython

SSLSOD

Logo

Self-Supervised Pretraining for RGB-D Salient Object Detection

Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, Xiang Ruan
⭐ arXiv » 🔥[Slide&极市平台推送]

The official repo of the AAAI 2022 paper, Self-Supervised Pretraining for RGB-D Salient Object Detection.

Saliency map

Google Drive / BaiduYunPan(d63j)

Trained Model

You can download all trained models at Google Drive / BaiduYunPan(0401).

Datasets

  • Google Drive / BaiduYunPan(1a4t)
  • We believe that using a large amount of RGB-D data for pre-training, we will get a super-strong SSL-based model even surpassing the ImageNet-based model. This survey of the RGB-D dataset may be helpful to you.

Training

  • SSL-based model
    1.Run train_stage1_pretext1.py
    2.Run get_contour.py (can generate the depth-contour maps for the stage2 training) 2.Load the pretext1 weights for Crossmodal_Autoendoer (model_stage1.py) and run train_stage2_pretext2.py
    3.Load the pretext1 and pretext2 weights for RGBD_sal (model_stage3.py) as initialization and run train_stage3_downstream.py
  • ImageNet-based model
    Set 'pretrained= Ture' for models.vgg16_bn(pretrained='True') in RGBD_sal (model_stage3.py) and run train_stage3_downstream.py

Testing

Run prediction_rgbd.py (can generate the predicted saliency maps)
Run test_score.py (can evaluate the predicted saliency maps in terms of fmax,fmean,wfm,sm,em,mae,mdice,miou,ber,acc).

BibTex

@inproceedings{SSLSOD,
  title={Self-Supervised Pretraining for RGB-D Salient Object Detection},
  author={Zhao, Xiaoqi and Pang, Youwei and Zhang, Lihe and  and Lu, Huchuan and Ruan, Xiang},
  booktitle={AAAI},
  year={2022}
}