Semantic Aware Attention Based Deep Object Co-segmentation

• Object Cosegmentation with Deep learning and Attention Module
• Capable of segment unseen obejct
• Cosegment multiple inputs with linear time.

Accepted by ACCV 2018 https://arxiv.org/abs/1810.06859

Co-segmentation result

Instant Group Co-segmentation result

Model

Model:

• Encoder : VGG16 (last layer before fc layer)
• Attention Module
• Decoder : Deconvolution layers

Preprocessed Dataset

You can download from here!

https://drive.google.com/open?id=1xHkdAoAPNG5Vixd3wMFZEtmLJ3rBQZSG

Pretrained model

Updating

Install

Here is the list of libraries you need to install to execute the code:

• python = 3.6
• pytorch = 0.4
• Pillow
• Numpy
• Glob

Train model

For example:

python train.py --train_data "Datasets/PascalVoc/image/" --train_label "Datasets/PascalVoc/colabel/train/" --train_txt "Datasets/PascalVoc/colabel/train.txt" --val_data "Datasets/PascalVoc/image/" --val_label "Datasets/PascalVoc/colabel/val/" --val_txt "Datasets/PascalVoc/colabel/val.txt" --model_path "model_path/"

Co-segment two images

For example:

python single_demo.py --image1 "demo/1.jpg" --image2 "demo/2.jpg" --output1 "demo/co_1.jpg"  --output2 "demo/co_2.jpg" --model "model_path/epoch1iter5000.pkl"

Co-segment multiple images

For example:

python group_demo.py --image_path "group_demo/images/" --output_path "group_demo/outputs/" --model "model_path/epoch1iter5000.pkl"

Inputs and outputs all seperate into two different directories.