HDFNet

(ECCV 2020) Hierarchical Dynamic Filtering Network for RGB-D Salient Object Detection

Author: Lart Pang(lartpang@163.com)

This is a complete, modular and easily modified code base based on PyTorch, which is suitable for the training and testing of significant target detection task model.

@inproceedings{HDFNet-ECCV2020,
    author = {Youwei Pang and Lihe Zhang and Xiaoqi Zhao and Huchuan Lu},
    title = {Hierarchical Dynamic Filtering Network for RGB-D Salient Object Detection},
    booktitle = ECCV,
    year = {2020}
}

NOTE:

In the link below, we provide the results of the two versions (with/without _STEREO).
Specifically, in the file with _STEREO, two versions of the STEREO dataset are provided.
- One of them contains 797 pictures, and the other contains 1000 pictures.
- In our paper, the results evaluated on the latter are used, which is exactly what is provided in the file without _STEREO.

[Results & PretrainedParams (j9qu)]

RGBD-DATASET
- https://github.com/jiwei0921/RGBD-SOD-datasets
- http://dpfan.net/d3netbenchmark/

Repository Details

backbone: Store some code for backbone networks.
loss: The code of the loss function.
module: The code of important modules.
network: The code of the network.
output: It saves all results.
utils: Some instrumental code.
- data/*py: Some files about creating the dataloader.
- transforms/*py: Some operations on data augmentation.
- metric.py: max/mean/weighted F-measure, S-measure, E-measure and MAE. (NOTE: If you find a problem in this part of the code, please notify me in time, thank you.)
- misc.py: Some useful utility functions.
- tensor_ops.py: Some operations about tensors.
config.py: Configuration file for model training and testing.
train.py: I think you can understand.
test.py and test.sh: These files can evaluate the performance of the model on the specified dataset. And the file test.sh is a simple example about how to configure and run test.py.

Usage

Environment

I provided conda environment configuration file (hdfnet.yaml), you can refer to the package version information.

And you can try conda env create -f hdfnet.yaml to create an environment to run our code.

Train your own model

Add your own module into the module.
Add your own network into the network and import your model in the network/__init__.py.
Modify config.py:
- change the dataset path: datasets_root
- change items in arg_config
  - model corresponds to the name of the model in network
  - suffix: finally, the form of <model>_<suffix> is used to form the alias of the model of this experiment and all files related to this experiment will be saved to the folder <model>_<suffix> in output folder
  - resume: set it to False to train normally
  - data_mode: set it to RGBD or RGB for using RGBD SOD datasets or RGB SOD datasets to train mdoel.
  - other items, like lr, batch_size and so on...
Run the script: python train.py

If the training process is interrupted, you can use the following strategy to resume the training process.

Set resume to True.
Run the script train.py again.

Evaluate model performance

There are two ways:

For models that have been trained, you can set resume to True and run the script train.py again.
Use the scripts test.sh and test.py. The specific method of use can be obtained by executing this command: python test.py --help.

Only evaluate generated predictions

You can use the toolkit released by us: https://github.com/lartpang/Py-SOD-VOS-EvalToolkit.

Related Works

(ECCV 2020 Oral) Suppress and Balance: A Simple Gated Network for Salient Object Detection: https://github.com/Xiaoqi-Zhao-DLUT/GateNet-RGB-Saliency
(ECCV 2020) A Single Stream Network for Robust and Real-time RGB-D Salient Object Detection: https://github.com/Xiaoqi-Zhao-DLUT/DANet-RGBD-Saliency
(CVPR 2020) Multi-scale Interactive Network for Salient Object Detection: https://github.com/lartpang/MINet

lmkzero/HDFNet