grassbro / ocor

Bi-directional Object-context Prioritization Learning for Saliency Ranking

OCOR

This is a reproduction of the CVPR'22 paper, Bi-directional Object-context Prioritization Learning for Saliency Ranking.

Installation

step 1: install pytorch and mmcv==1.3.9 referring to this.

step 2: clone this repository and execute:

python setup.py develop

step 3: install apex following this. (optional)

Dataset

Download COCO-style JSON files for the ASSR dataset from: https://pan.baidu.com/s/1XvYwBCn3sc6lAlJbJ94gUQ (pwd: ocor)

Training

bash tools/dist_train.sh configs/ocor/ocor_swin...py num_gpus 

Inference & Evaluation

python inference.py

python evaluate_SOR.py

Pre-trained Model

Download it from our Baidu cloud: https://pan.baidu.com/s/15tINLiVC8kPQm6xqxyaJlA (pwd: ocor), then use it for fine-tuning, and inference.

Results

Visual results for ASSR dataset: https://pan.baidu.com/s/1V3MzSBWI_5UNSfC1YCVf3g (pwd: ocor)

Citation

@inproceedings{tian2022bi,
  title={Bi-directional object-context prioritization learning for saliency ranking},
  author={Tian, Xin and Xu, Ke and Yang, Xin and Du, Lin and Yin, Baocai and Lau, Rynson WH},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5882--5891},
  year={2022}
}

ocor's Issues

训练代码

作者你好，请问dist_train.sh或者训练代码在哪里~谢谢

saliency ranking loss

作者您好，论文中提到使用pair-wise ranking loss来训练排序头，但是我在代码中并没有找到，烦请您告知位置。

ocor module以及saliency ranking loss

作者你好，在看您的代码时，我没有找到您论文中提到的ocor module以及训练中用到的saliency ranking loss，请问可以告知在哪嘛？

预训练模型

作者你好，请问开源的预训练模型是在什么数据集上预训练的