crescent-ao / mpanet

This is the official implementation of the MPANet

This is no final version.I will commit all the changes after I finished my experiment. If you need it, please open issue so that I can give you the newest code. Thanks

MPANet:Multi-Patch Attention for Infrared Target Detection

Hello,every one.This is MPAnet offical implementation.In this repo ,I will integrate nearly all metric to this model.Because it is my first one to update a complete repo. It exists some errors in this repo.If you have any problems, I am able to try my best to resolve your problems.

MPANet network

A MPANet framework based on the axial-attention encoder and the multi-scale patch branch (MSPB) structure is proposed to highlight the small targets and suppress background without any classification backbones. In the designed MSPB, coarse-grained features extracted by the large-scale branch and fine-grained features extracted by the local branches are fused through the de- veloped bottom-up structure. Extensive experiments on the public SIRST dataset. The segmentation results are as follows:

✨✨🎉🎉

MPANet has been accepted by IGARSS 2022 (oral).I will release a new dataset which consist of 1077 images. It includes sirst dataset, MD vs FA dataset and completely new infrared data with high quality annotations.In the meanwhile, I alse make some improvements on MPANet, which make its inference speed faster 60% than the older in a little cost. It overcomes the problem of inability to communicate between MPANet sub-patches through a pyramid-like patch fusion method.

mpanet's Issues

对比模型请教

四月伊始, 在这里预祝您有一个愉快的四月^_^.

我叫陈壮, 目前同样正在研究红外小目标检测的相关问题, 看到您公开的MPANet后我感到十分振奋, 这真是一项极具创造性的工作, 祝您论文投稿顺利.

拜读您的工作后, 我相信您肯定要对比ALCNet这一模型, 这一定是我们对比模型中一座绕不过的"大山".

可是这个模型的作者戴博士习惯使用MXNET, 这对于习惯Pytorch的我来说十分煎熬, 尝试转化为Pytorch版本后也是问题百出.

因此冒昧来信, 希望您能够开源您实现的Pytorch版本的ALCNet, 我相信这对于整个研究领域内的人来说都是一件值得喝彩的壮举.

再次感谢您在红外小目标目标领域内的工作, 以及您热爱&&共享的开源精神，这为同样研究这个问题的我们提供了宝贵的思路与优质数据集.

祝学业顺利 生活愉快.

陈壮