apcc-geoslegend / u2net-tensorflow

Tensorflow implementation of U^2-Net [...] for Salient Object Detection

u2net-tensorflow

A tensorflow implementation of the U2-Net: Going Deeper with Nested U-Structure for Salient Object Detection using Keras & Functional API

Based on the PyTorch version by NathanUA, PDillis, vincentzhang, and chenyangh

Notebook

If you just want to play with the model, I've setup a Google Colab Notebook that lets you train the model on DUTS-TR, and it's fun to watch the model learn to mask an image of the Space Needle that it's never seen before while it trains. Training takes ~60 minutes to get to noticeable results, but you should train for several hours to use it for testing. Please let me know if you have any questions.

Setup

virtualenv venv
source venv/bin/activate
pip install tensorflow matplotlib opencv-python wget

Training

OPTIONAL: Download the DUTS-TR dataset and unzip into the data directory to load the training set:

wget http://saliencydetection.net/duts/download/DUTS-TR.zip
unzip DUTS-TR.zip -d data

If train.py does not detect the dataset is present when run, it will automatically try to download and setup the dataset before initiating training. If you have a custom dataset, you can update dataset_dir, image_dir, and mask_dir in config.py. Images will be rescaled

To begin training, simply run:

python train.py

Weights are automatically saved every save_interval iterations to weights/u2net.h5. These can be overwritten by passing the appropriate arguments. See python train.py -h for args.

Evaluation

Use eval.py to evaluate the model on images:

python eval.py --weights=weights/u2net.h5 --image=examples/skateboard.jpg

By default, the output images are saved in the out subdirectory

Custom Usage

The U2NET class can be used to instatiate a modular instance of the U2-Net network

from model.u2net import U2NET

u2net = U2NET()
out = u2net(inp)