SODBASNET 0.0.5

BASNET model created using tensorflow.

SOD-BASNET-Py

BASNet: Boundary-Aware Salient Object Detection

• Deep Convolutional Neural Networks have been adopted for salient object detection and achieved the state-of-the-art performance. Most of the previous works however focus on region accuracy but not on the boundary quality.
• In this project, we propose a predict-refine architecture, BASNet, and a new hybrid loss for Boundary-Aware Salient object detection. Specifically, the architecture is composed of a densely supervised Encoder-Decoder network and a residual refinement module, which are respectively in charge of saliency prediction and saliency map refinement.
• The Research paper referred is BASNet: Boundary-Aware Salient Object Detection

Objective

• The Official Code for the Basnet Model is provided by the authors on the github.
• The Official Deep Learning model is made using Pytorch , my Goal is to create a similar model using tensorflow.
• Why ? Whilst Pytorch provide better developer experience and error handeling, I find Tensorflow to be a great ML Framework to work with for Beginners .
• So the objective for this repo is to create Basnet Model using Tensorflow.

Architecture

• The Architecture Proposed by the authors is predict-refine architecture.
• The Author have used Transfer Learning to improve model performance . They Used First 4 layers of the resnet34 model which is pretrained in imagenet dataset.
• But for the sake of understanding abstract architecture of the Basnet model I haven't used the transfer learning method , instead created each layer individually.

Loss

• The hybrid loss guides the network to learn the transformation between the input image and the ground truth in a three-level hierarchy – pixel-, patch- and map- level – by fusing Binary Cross Entropy (BCE), Structural SIMilarity (SSIM) and Intersectionover- Union (IoU) losses.

Thank You

• This is my attempt to build the Basnet model using the research paper . If you find any error or mistakes let me know . Love all the feedbacks.