FishNet: A Versatile Backbone for Image, Region, and Pixel Level Prediction

By Shuyang Sun et al

Published on Jan. 11, 2019

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Identity Mappings in Deep Residual Networks and Isolated Convolution
3. The FishNet
4. Feature Refinement
5. Related Works

Summary

FishNet is a network structure designed to unify the advantages of networks for pixel-level, region-level, and image-level tasks. It addresses the problem of propagating gradients from deep layers to shallow layers. The network consists of three parts: tail, body, and head, each serving different purposes in preserving and refining features. FishNet demonstrates improved performance in image classification and tasks like object detection and segmentation. The paper discusses the design principles and mechanisms of FishNet, highlighting its effectiveness in handling features of varying resolutions and depths.