Awesome

Saliency-Ranking

Code release for the TPAMI 2021 paper "Instance-Level Relative Saliency Ranking with Graph Reasoning" by Nian Liu, Long Li, Wangbo Zhao, Junwei Han, and Ling Shao.

Installation

See INSTALL.md.

Data Preparation

Download the datatset from Baidu Driver (rx96) or Google Driver and unzip them to './dataset'. Then the structure of the './dataset' folder will show as following:

-- dataset
   |-- Annotations
   |   |-- | train.pkl
   |   |-- | test.pkl
   |-- Images
   |   |-- train
   |   |-- |-- | rgb
   |   |-- |-- |-- | COCO_train2014_000000000110.jpg ...
   |   |-- |-- | gt
   |   |-- |-- |-- | COCO_train2014_000000000110.png ...
   |   |-- test
   |   |-- |-- | rgb
   |   |-- |-- |-- | COCO_val2014_000000000192.jpg ...
   |   |-- |-- | gt
   |   |-- |-- |-- | COCO_val2014_000000000192.png ...

Training model

1. Download the pretrained model (our modified Mask R-CNN model for salient instance segmentation) from Baidu Driver(spq9) or Google Driver and put it into ./model folder.
2. Run python ./tool/plain_train_net.py.
3. The trained model will be saved in ./output folder. Additionally, the evaluaion results produced during training process will be saved in ./output/SA_SOR.txt and ./output/MAE.txt.

Testing model

1. Download our trained model from Baidu Driver(fhz7) or Google Driver. Rename it as 'final_model.pth' and put it into ./model folder.
2. Run python ./tool/plain_test_net.py.
3. The prediction images will be saved in ./prediction. And the metric score, SA_SOR and MAE, will be printed.

Result

The prediction results of our dataset can be download from prediction (k1jr).

Metric

We propose a new evaluation metirc for this task, which comprehensively considers salient instance detection, segmentation, and ranking performance. See SA-SOR.

We test the SOR metric of our model by first predicting the saliency maps via the test.py in our code and then feeding these saliency maps to the official SOR code (a matlab code can be download from SOR code(dems)).