LFSOD-CDINet

This project provides the code and results for 'Light Field Salient Object Detection with Sparse Views via Complementary and Discriminative Interaction Network', IEEE TCSVT, 2023. paper link

Network architecture

Requirements

python 3.7 + TensorFlow 1.14.0

Note: We provide a modified layer.py (code: 7d8i) for TensorFlow 1.14.0. The added layer_norm_initialized() enables initializing Layer_Norm with pre-trained parameters. You can put it under 'your_Anaconda_envs/Lib/site-packages/tensorflow/contrib/layers/python/layers/' folder.

Training

1. Download the TrainingSet (code: t7gt) and put it under './dataset/' folder.
2. Download the pre-trained vgg-16 model (code: kq1o) and mpi model (code: c3tj) and put them under './models/' folder.
3. Run train.py (default to the HFUT-Lytro Illum dataset).

Test using pre-trained model

1. Download the TestSet (code: hdl2) and put it under './dataset/' folder.
2. Download our pre-trained model_HFUT (code: k28i) and model_DUTLF-V2 (code: h8ou) and put them under './checkpoints/' folder.
3. Run test.py. The SOD results will be saved under './results/' folder.

Note: In the paper, we use model_HFUT to test the HFUT-Lytro Illum & HFUT-Lytro datasets and use model_DUTLF-V2 to test the DUTLF-V2 dataset.

Saliency maps and performance

We provide results (code: lau2) of our CDINet on 3 datasets (HFUT-Lytro Illum, HFUT-Lytro and DUTLF-V2)

Citation

@ARTICLE{LFSOD-CDINet,  
  title={Light Field Salient Object Detection with Sparse Views via Complementary and Discriminative Interaction Network},
  author={Yilei Chen and Gongyang Li and Ping An and Zhi Liu and Xinpeng Huang and Qiang Wu},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2024},
  volume={34},
  number={2},
  pages={1070-1085},
  month={Feb.}}            

Any questions regarding this work can contact yileichen@shu.edu.cn.