HQ_SOD_repo

High-Quality Saliency Object Detection repository using FCN and DKN, by Avia Israeli and Omri Levi (TAU, Deep Learning by Raja Giryes, 0510-7255, Fall 2019).

Project full report can be found here: https://drive.google.com/file/d/1AEF89GNBrMD5KdLUGSAbhaeC0AqNyRVS/view?usp=sharing

1. Introduction:

This repository includes the code for training and testing the high-quality, yet light-weight, SOD model described in the paper. It consist of main models:

• LRP [1] - Low-Resolution Predictor module.
• DKN [2] - Deformable Kernel Networks - the joint-filtering refinement module.
• Full SOD module. This repository is for personal educational use only.

2. Dataset:

Download DUTS [3] dataset from here: http://saliencydetection.net/duts/ The dataset should contain 2 folders: DUTS-TR, DUTS-TE for training and testing. Every folder should contain 2 sub-folders: DUTS-TR-Image and DUTS-TR-Mask for training (or DUTS-TE-Image and DUTS-TE-Mask for testing). In order to use the default path (as defined in our code) put the dataset in a different folder from the repository named "datasets" or use your own path.

3. Pre-trained weights:

Download pre-trained weights from here: https://drive.google.com/open?id=1sRoEUvwg5CLxHigHnx4I1-XADsSQuXny Extract the files to "HQ_SOD_repo/saved_models/" (DKN parameters are already included in the repository in HQ_SOD_repo/DKN/saved_model or HQ_SOD_repo/DKN/parameter).

4. Requirements:

Use the yml file (HQ_SOD_repo/environment_requirements.yml) to see the requirements (we recommend setting the environment via anaconda).

5. How to run:

Every model can be either trained or tested. The training process can begin from scratch or from pre-trained parameters. For training, change the relevant variables in the parser or use the defaults. For testing, change the relevant variables according to your parameters and inputs paths.

5.1. LRP

- Inference: run HQ_SOD_repo/LRP/train_unet_upsampled_model.py - change the parameter "--b_train" to False and make sure the path to the pre-trained parameters in "--parameter" is correct
- Train: run HQ_SOD_repo/LRP/train_unet_upsampled_model.py - change the parameter "--b_train" to True. For training from scratch use "--parameter" value "None".

5.2. DKN

- Inference: run HQ_SOD_repo/DKN/dkn_inference.py - change the "--rgb_folder" and "--mask_folder" for the correct paths (or use defaults), change image name "--image_name"
- Train: run HQ_SOD_repo/DKN/dkn_train.py - make sure "--train_guide_dir" and "--test_guide_dir" are correct.

5.3. Full SOD model

- Inference: run HQ_SOD_repo/inference_full_sod_model.py - make a list of image names in "input_images" and make sure the paths input image directory and output image directory in "--rgb_folder" and "--mask_folder" are correct. Make sure the path for the pre-trained parameters in "--merged_model_params" is correct.
- Train: run HQ_SOD_repo/train_full_sod_model.py - make sure "--train_input_images_dir" and "--test_input_images_dir" are correct. Make sure the path for the pre-trained parameters in "--pre_trained_full_network_params" is correct.

6. Credits and References:

• [1] - Inspired by https://github.com/ternaus/TernausNet/blob/master/unet_models.py
• [2] - DKN Github pytorch - https://github.com/jun0kim/DKN
• [3] - DUTS dataset - http://saliencydetection.net/duts/ - Lijun Wang, Huchuan Lu, Yifan Wang ,Mengyang Feng, Dong Wang, Baocai Yin, Xiang Ruan, "Learning to Detect Salient Objects with Image-level Supervision", CVPR2017