1. 📎 Paper Link
2. 📃 Requirements
3. 📂 Toy Figure-Ground Dataset
4. ✏️ Usage
5. 📊 Results
6. ✉️ Statement
7. 🔍 Citation

Exploring Figure-Ground Assignment Mechanism in Perceptual Organization

Authors: Wei Zhai, Yang Cao, Jing Zhang, & Zheng-Jun Zha,

• Python == 3.7
• Pytorch == 1.7.0 (pytorch.org)
• Torchvision == 0.5.0
• Cuda ==11.0 (cuda)

Need to download:

• Pascal VOC dataset (http://host.robots.ox.ac.uk/pascal/VOC/)
• Describable Textures Dataset (https://www.robots.ox.ac.uk/~vgg/data/dtd/)

Create Easy Set:

python extractor.py --level E --PascalDir ${Pascal Dir} --DTDDir ${DTD Dir} --SaveDir ${Save Dir}

Create Normal Set:

python extractor.py --level N --PascalDir ${Pascal Dir} --DTDDir ${DTD Dir} --SaveDir ${Save Dir}

Create Hard Set:

python extractor.py --level H --PascalDir ${Pascal Dir} --DTDDir ${DTD Dir} --SaveDir ${Save Dir}

To train a model on Easy Set, run Train.py with the desired model:

python Train.py --batch_size_train 16 --accumulation_steps 1 --check_ite 6000 --lr 0.0001 --step_size 50 --begin_ite 6000 --noFG 1 --level E

To train a model on Normal Set, run Train.py with the desired model:

python Train.py --batch_size_train 16 --accumulation_steps 1 --check_ite 6000 --lr 0.0001 --step_size 50 --begin_ite 6000 --noFG 1 --level N

To train a model on Hard Set, run Train.py with the desired model:

python Train.py --batch_size_train 16 --accumulation_steps 1 --check_ite 6000 --lr 0.0001 --step_size 50 --begin_ite 6000 --noFG 1 --level H

To test a model, run Test.py with the desired model on different datasets:

python Test.py --modelname ${model_name} --batchsize 4

FGA-Net has a better capability of Figure-Ground assignment. (a) FGA-Net outperforms the representative models for Figure-Ground Segregation on all three datasets. Different dash boxes represent different levels. The x-axis and the y-axis are the S-measure and IoU scores, respectively. UNet+E/DeepLabv3+E indicates UNet/DeepLabv3 with auxiliary edge supervision. The numerical results are detailed in Supp. Material. (b) FGA-Net is more parameter-efficient than representative models for the Figure-Ground Segregation test. The x-axis shows the computational complexity, while the y-axis depicts the IoU performance. (c) FGA-Net is more data-efficient than representative models. The x-axis and the y-axis represent different proportions of the normal dataset and the IoU score, respectively. (d) Visual comparisons on the Normal dataset.

This project is for research purpose only, please contact us for the licence of commercial use. For any other questions please contact wzhai056@mail.ustc.edu.cn.

none