Revisiting Image Pyramid Structure for High Resolution Salient Object Detection

October 31, 2022 · View on GitHub

Salient Object detection - examples created using published model

Newsletter describing this paper

App using the model described in this paper

Revisiting Image Pyramid Structure for High Resolution Salient Object Detection

PyTorch implementation of Revisiting Image Pyramid Structure for High Resolution Salient Object Detection (InSPyReNet)

To appear in the 16th Asian Conference on Computer Vision (ACCV2022)

Abstract: Salient object detection (SOD) has been in the spotlight recently, yet has been studied less for high-resolution (HR) images. Unfortunately, HR images and their pixel-level annotations are certainly more labor-intensive and time-consuming compared to low-resolution (LR) images. Therefore, we propose an image pyramid-based SOD framework, Inverse Saliency Pyramid Reconstruction Network (InSPyReNet), for HR prediction without any of HR datasets. We design InSPyReNet to produce a strict image pyramid structure of saliency map, which enables to ensemble multiple results with pyramid-based image blending. For HR prediction, we design a pyramid blending method which synthesizes two different image pyramids from a pair of LR and HR scale from the same image to overcome effective receptive field (ERF) discrepancy. Our extensive evaluation on public LR and HR SOD benchmarks demonstrates that InSPyReNet surpasses the State-of-the-Art (SotA) methods on various SOD metrics and boundary accuracy.

Demo :rocket:

Demo1	Demo2
<img src=./figures/fig_demo1.gif height="350px" width="350px">	<img src=./figures/fig_demo2.gif height="350px" width="350px">

Architecture

InSPyReNet	pyramid blending
<img src=./figures/fig_architecture.png height="350px" width="350px">	<img src=./figures/fig_pyramid_blending.png height="350px" width="350px">

1. Create environment

Create conda environment with following command conda create -y -n inspyrenet python=3.8
Activate environment with following command conda activate inspyrenet
Install requirements with following command pip install -r requirements.txt

2. Preparation

Item	Destination Folder	OneDrive	GDrive
Train Datasets	`data/Train_Dataset/...`	Link	Link
Test Datasets	`data/Test_Dataset/...`	Link	Link
Res2Net50 checkpoint	`data/backbone_ckpt/*.pth`	Link	Link
SwinB checkpoint	`data/backbone_ckpt/*.pth`	Link	Link

3. Train & Evaluate

Train InSPyReNet (SwinB)

python run/Train.py --config configs/InSPyReNet_SwinB.yaml --verbose

Train with extra training datasets can be done by just changing Train.Dataset.sets in the yaml config file, which is just simply adding more directories (e.g., HRSOD-TR, HRSOD-TR-LR, UHRSD-TR, ...):

Train:
  Dataset:
      type: "RGB_Dataset"
      root: "data/RGB_Dataset/Train_Dataset"
      sets: ['DUTS-TR'] --> ['DUTS-TR', 'HRSOD-TR-LR', 'UHRSD-TR-LR']

Inference for test benchmarks

python run/Test.py --config configs/InSPyReNet_SwinB.yaml --verbose

Evaluate metrics

python run/Eval.py --config configs/InSPyReNet_SwinB.yaml --verbose

Please note that we only uploaded the low-resolution (LR) version of HRSOD and UHRSD due to their large image resolution. In order to use them, please download them from the original repositories (see references below), and change the directory names as we did to the LR versions.

4. Inference on your own data

You can inference your own single image or images (.jpg, .jpeg, and .png are supported), single video or videos (.mp4, .mov, and .avi are supported), and webcam input (ubuntu and macos are tested so far).
python run/Inference.py --config configs/InSPyReNet_SwinB.yaml --source [SOURCE] --dest [DEST] --type [TYPE] --gpu --jit --verbose
- SOURCE: Specify your data in this argument.
  - Single image - image.png
  - Folder containing images - path/to/img/folder
  - Single video - video.mp4
  - Folder containing videos - path/to/vid/folder
  - Webcam input: 0 (may vary depends on your device.)
- DEST (optional): Specify your destination folder. If not specified, it will be saved in results folder.
- TYPE: Choose between map, green, rgba, blur
  - map will output saliency map only.
  - green will change the background with green screen.
  - rgba will generate RGBA output regarding saliency score as an alpha map. Note that this will not work for video and webcam input.
  - blur will blur the background.
- --gpu: Use this argument if you want to use GPU.
- --jit: Slightly improves inference speed when used.
- --verbose: Use when you want to visualize progress.

5. Checkpoints

Note: If you want to try our trained checkpoints below, please make sure to locate latest.pth file to the Test.Checkpoint.checkpoint_dir.

Trained with LR dataset only (DUTS-TR, 384 X 384)

Backbone	Train DB	Config	OneDrive	GDrive
Res2Net50	DUTS-TR	InSPyReNet_Res2Net50	Link	Link
SwinB	DUTS-TR	InSPyReNet_SwinB.yaml	Link	Link

Trained with LR+HR dataset (with LR scale 384 X 384)

Backbone	Train DB	Config	OneDrive	GDrive
SwinB	DUTS-TR, HRSOD-TR-LR	InSPyReNet_SwinB.yaml	Link	Link
SwinB	HRSOD-TR-LR, UHRSD-TR-LR	InSPyReNet_SwinB.yaml	Link	Link
SwinB	DUTS-TR, HRSOD-TR-LR, UHRSD-TR-LR	InSPyReNet_SwinB.yaml	Link	Link

*-LR denotes resized into low-resolution scale (i.e., 384 X 384).

Trained with LR+HR dataset (with HR scale 1024 X 1024)

Backbone	Train DB	Config	OneDrive	GDrive
SwinB	DUTS-TR, HRSOD-TR	InSPyReNet_SwinB.yaml	Link	Link
SwinB	HRSOD-TR, UHRSD-TR	InSPyReNet_SwinB.yaml	Link	Link

Trained with Massive SOD Datasets (with LR scale 384 x 384, Not in the paper, just for fun!)

Backbone	Train DB	Config	OneDrive	GDrive
SwinB	DUTS-TR, DUTS-TE, FSS-1000, MSRA-10K, ECSSD, HRSOD-TR-LR, UHRSD-TR-LR	InSPyReNet_SwinB.yaml	Link	Link

6. Pre-Computed Saliency Maps

Note: Due to the cloud memory shortage, we only provide results trained on DUTS-TR only. Please generate yourself for the models with extra training datasets if you need.

Backbone	DUTS-TE	DUT-OMRON	ECSSD	HKU-IS	PASCAL-S	DAVIS-S	HRSOD-TE	UHRSD-TE
Res2Net50	Link	Link	Link	Link	Link	N/A	N/A	N/A
SwinB	Link	Link	Link	Link	Link	Link	Link	Link

7. Results

Quantitative

LR Benchmark	HR Benchmark	HR Benchmark (Trained with extra DB)
<img src=./figures/fig_quantitative.png height="250px" width="250px">	<img src=./figures/fig_quantitative2.png height="250px" width="250px">	<img src=./figures/fig_quantitative3.png height="250px" width="250px">

Qualitative

DAVIS-S & HRSOD	UHRSD	UHRSD (Trained with extra DB)
<img src=./figures/fig_qualitative.png height="250px" width="250px">	<img src=./figures/fig_qualitative2.png height="250px" width="250px">	<img src=./figures/fig_qualitative3.jpg height="250px" width="250px">

Citation

@article{kim2022revisiting,
  title={Revisiting Image Pyramid Structure for High Resolution Salient Object Detection},
  author={Kim, Taehun and Kim, Kunhee and Lee, Joonyeong and Cha, Dongmin and Lee, Jiho and Kim, Daijin},
  journal={arXiv preprint arXiv:2209.09475},
  year={2022}
}

References

Backbones: Res2Net, Swin Transformer
Datasets:
- LR Benchmarks: DUTS, DUT-OMRON, ECSSD, HKU-IS, PASCAL-S
- HR Benchmarks: DAVIS-S, HRSOD, UHRSD
Evaluation Toolkit
- SOD Metrics (e.g., S-measure): PySOD Metrics
- Boundary Metric (mBA): CascadePSP