Home

Awesome

SODGAN

This is the official code and data release for:

Synthetic Data Supervised Salient Object Detection, Accepted by ACM Multimedia 2022

Requirements

Training

To reproduce paper **Synthetic Data Supervised Salient Object Detection via **:

cd SODGAN
  1. Run Step1: training mask generator.
  2. Run Step2: synthesizing annotation-image pairs.
  3. Run Step3: Train SOD model.

1. Training Mask Generator

we take training stlyegan as an example:

a. Download pretrained model from StyleGAN [https://github.com/NVlabs/stylegan]. Put pretrained model in 'your/path/' and you have revised the path of 'stylegan_checkpoint' of ./experiments/cat_sod.json

b. Download Dataset from [https://pan.baidu.com/s/1e7SRXVTqTxR3CQJEtq_HFg] (fetch code:2nab ). Unzip stylegan datasets into './data/'.You have to revise 'annotation_mask_path', 'testing_path', 'average_latent' of ./experiments/cat_sod.json

c. python train_stylegan_G_mask.py --exp experiments/stylegan/cat_sod.json  --test False

2. Synthesizing annotation-image pairs

python train_stylegan_G_mask.py --exp experiments/stylegan/cat_sod.json  --test True  --resume [your trained model path]

Example of sampling images and annotation:

<img src = "./figures/stylegan.jpg" width="80%"/>

or

python train_biggan_G_mask.py --exp experiments/biggan/all.json  --test True  --resume [your trained model path]

Example of sampling images and annotation:

<img src = "./figures/biggan.jpg" width="90%"/>

3. Train SOD model

These synthesized data can be used for training off-the-shelf saliency networks.

Pretrained Model

You can skip the step 1 and use our pretrained model as below: \

Mask Generator (BigGAN) [https://pan.baidu.com/s/1Nr1OfQq7d_6hakDo8z218A] (fetch code:lb6u )
Mask Generator (StyleGAN cat) [https://pan.baidu.com/s/1_yhbGVzH92BEU8P66RtLwg] (fetch code: pkw8 )

Saliency maps

We also provide saliency maps for comparisons [https://pan.baidu.com/s/1WN613RbPeSzmZiISMymt_Q] (fetch code:b818 )

Comparison with state-of-the-art

<img src = "./figures/table1.jpg" width="80%"/>

License

For any code dependency related to StyleGAN, StyleGAN2, and BigGAN, the license is under the Creative Commons BY-NC 4.0 license by NVIDIA Corporation. To view a copy of this license, visit LICENSE.