This is a PyTorch implementation of LO-Det, YOLOv3+MobileNetv2, and YOLOv4.
如果觉得万一有点用，欢迎引用我们的论文。谢谢。
@ARTICLE{9390310,
author={Z. {Huang} and W. {Li} and X. -G. {Xia} and H. {Wang} and F. {Jie} and R. {Tao}},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={LO-Det: Lightweight Oriented Object Detection in Remote Sensing Images},
year={2021},
volume={},
number={},
pages={1-15},
doi={10.1109/TGRS.2021.3067470}}
##10.9 更新预告 国庆弄了个知识蒸馏和量化压缩的版本，DOTA1.0_OBB目前mAP能到71.38，和我5月那个GGHL的工作持平了，丢了个会议，等审稿完代码会和v2版本一起放出来。 另外根据大家都建议修改了一些bug和做了一些优化，有问题的可以看issues或者邮件联系我。谢谢。
##一些解释和惯例的碎碎念
代码这几天会分批陆续上传，因为有一些函数我重新命名和封装了（因为代码是基于更复杂的检测框架的，要从其中把LO-Det的部分单独拆出来）。模型权重字典的key我也需要重新转换一下。
目前是一个粗糙版本，因为想着尽快开源让大家用，还没来得及重构和优化，可能会有一些bug，遇到任何问题都可以在issues里面留言或者给我发邮件（邮件地址论文里有，一作那个就是），我会尽最大努力帮你们跑通实验和复现结果。
这个代码和我的另一个工作NPMMR-Det用的是一个框架，大部分基础函数和用法都是一样的，有不明白的也可以看那个仓库。后续（有时间的话）会在那边仓库出教程和每行代码的注释。
关于在嵌入式设备上实验以及Nvidia TX2\Xavier的其他使用问题也欢迎交流。
最近事儿挺多，代码优化请允许我延迟一些时间（大概一个月之内），我会尽快完成。
有计划出一版LO-Detv2的魔改版，正在做实验。这个做完以后接下来的目标检测工作可能会逐渐淡化遥感方向了，对各方面都有点失望吧。
The High-precision version NPMMR-Det
A Novel Nonlocal-aware Pyramid and Multiscale Multitask Refinement Detector for Object Detection in Remote Sensing Images
@ARTICLE{9364888,
author={Z. {Huang} and W. {Li} and X. -G. {Xia} and X. {Wu} and Z. {Cai} and R. {Tao}},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={A Novel Nonlocal-Aware Pyramid and Multiscale Multitask Refinement Detector for Object Detection in Remote Sensing Images},
year={2021},
volume={},
number={},
pages={1-20},
doi={10.1109/TGRS.2021.3059450}}
Clone不Star，都是耍流氓~
Linux (Ubuntu 18.04, GCC>=5.4) & Windows (Win10, VS2019)
CUDA 11.1, Cudnn 8.0.4
-
For RTX20/Titan RTX/V100 GPUs （I have tested it on RTX2080Ti, Titan RTX, and Tesla V100 (16GB)）
cudatoolkit==10.0.130
numpy==1.17.3
opencv-python==3.4.2
pytorch==1.2.0
torchvision==0.4.0
pycocotools==2.0 (In the ./lib folder)
dcnv2==0.1 (In the ./lib folder)
...
The installation of other libraries can be carried out according to the prompts of pip/conda -
For RTX30 GPUs （I have tested it on RTX3080 and RTX3090 GPUs）
cudatoolkit==11.0.221
numpy==1.17.5
opencv-python==4.4.0.46
pytorch==1.7.0
torchvision==0.8.1
pycocotools==2.0 (In the ./lib folder)
dcnv2==0.1 (In the ./lib folder)
...
- git clone this repository
- Install the libraries in the ./lib folder
(1) DCNv2
cd ./NPMMR-Det/lib/DCNv2/
sh make.sh
(2) pycocotools
cd ./NPMMR-Det/lib/cocoapi/PythonAPI/
sh make.sh
- DOTA dataset and its devkit
- DIOR dataset, password: 554e
(1) VOC Format
You need to write a script to convert them into the train.txt file required by this repository and put them in the ./data folder.
For the specific format of the train.txt file, see the example in the /data folder.
(2) MSCOCO Format
put the .json file in the ./data folder.
- train
python train.py - test
python test.py
Modify ./cfg/cfg_npmmr.py, please refer to the comments in this file for details
（https://drive.google.com/drive/folders/1cj0IKZ6rSVQRnt27AWiae0mSKhlLRQgs?usp=share_link）
If you have any questions, please ask in issues.
If you find bugs, please let me know and I will debug them in time. Thank you.
I will do my best to help you run this program successfully and get results close to those reported in the paper.
In addition to YOLOv3, YOLOv4 has also been initially implemented in this repository.
Some of the plug-and-play modules (many many Attentions, DGC, DynamicConv, PSPModule, SematicEmbbedBlock...) proposed in the latest papers are also collected in the ./model/plugandplay, you can use them and evaluate their performance freely. If it works well, please share your results here. Thank you.
(1) Model Pruning
(2) ONNX & TensorRT
...
https://github.com/Shank2358/NPMMR-Det
https://github.com/Peterisfar/YOLOV3
https://github.com/argusswift/YOLOv4-pytorch
https://github.com/ultralytics/yolov5
https://github.com/pprp/SimpleCVReproduction
This project is released under the Apache 2.0 license.