Awesome

OBGNet

Code for ACM MM 2021 paper "Occlusion-aware Bi-directional Guided Network for Light Field Salient Object Detection"

Overall Architecture

Datasets

We train and evaluate our model on DUTLF-v2 dataset.

Requirements

• Ubuntu 18.04
• torch 1.7.1
• python 3.8
• opencv-python 4.5.3.56
• imageio 2.4.1

Train

We train our model on multiple GPUs.

If you want to retrain our model, the process is as follows:

1. Please adjust hyperparameters in 'train_multiGPUs.py' , such as 'root', 'batch_size', and so on.

2. Please set CUDA devices in train_start.sh.

3. cd to the code path, then start training.

   nohup sh train_start.sh > log/xxx.txt 2>&1 &

Please note that the 'batch_size' should be greater than or equal to 16 for model generalization.

Test

You can download the checkpoint we provided (Baidu Pan, code:uk24).

Adjust paths in 'test.py' and run it to obtain predictions.

We also provide results of our model (Baidu Pan, code:dy78).

Contact us

If you have any questions, please contact us (20120370@bjtu.edu.cn).