TRACER: Extreme Attention Guided Salient Object Tracing Network

This paper was accepted at AAAI 2022 SA poster session. [pdf]

Updates

[09/06/2022] Demo has been released on Try it now!

[06/17/2022] Now, fast inference mode offers a salient object result with the mask.
We have improved a result quality of salient object as follows.
You can get the more clear salient object by tuning the threshold. We will release initializing TRACER with a version of pre-trained TE-x.

[04/20/2022] We update a pipeline for custom dataset inference w/o measuring.

• Run main.py scripts.

  
  TRACER
  ├── data
  │   ├── custom_dataset
  │   │   ├── sample_image1.png
  │   │   ├── sample_image2.png
    .
    .
    .

python main.py inference --dataset custom_dataset/ --arch 7 --img_size 640 --save_map True

Datasets

All datasets are available in public.

• Download the DUTS-TR and DUTS-TE from Here
• Download the DUT-OMRON from Here
• Download the HKU-IS from Here
• Download the ECSSD from Here
• Download the PASCAL-S from Here
• Download the edge GT from Here.

Data structure

TRACER
├── data
│   ├── DUTS
│   │   ├── Train
│   │   │   ├── images
│   │   │   ├── masks
│   │   │   ├── edges
│   │   ├── Test
│   │   │   ├── images
│   │   │   ├── masks
│   ├── DUT-O
│   │   ├── Test
│   │   │   ├── images
│   │   │   ├── masks
│   ├── HKU-IS
│   │   ├── Test
│   │   │   ├── images
│   │   │   ├── masks
      .
      .
      .

Requirements

• Python >= 3.7.x
• Pytorch >= 1.8.0
• albumentations >= 0.5.1
• tqdm >=4.54.0
• scikit-learn >= 0.23.2

Run

• Run main.py scripts.

  
  # For training TRACER-TE0 (e.g.)
  python main.py train --arch 0 --img_size 320

python main.py test --exp_num 0 --arch 0 --img_size 320

• Pre-trained models of TRACER are available at here
• Change the model name as 'best_model.pth' and put the weights to the path 'results/DUTS/TEx_0/best_model.pth'
  (here, the x means the model scale e.g., 0 to 7).
• Input image sizes for each model are listed belows.

Configurations

--arch: EfficientNet backbone scale: TE0 to TE7.
--frequency_radius: High-pass filter radius in the MEAM.
--gamma: channel confidence ratio \gamma in the UAM.
--denoise: Denoising ratio d in the OAM.
--RFB_aggregated_channel: # of channels in receptive field blocks.
--multi_gpu: Multi-GPU learning options.
--img_size: Input image resolution.
--save_map: Options saving predicted mask.

Citation

@article{lee2021tracer,
  title={TRACER: Extreme Attention Guided Salient Object Tracing Network},
  author={Lee, Min Seok and Shin, WooSeok and Han, Sung Won},
  journal={arXiv preprint arXiv:2112.07380},
  year={2021}
}