zhpmatrix / visdrone2018

ECCV2018(Challenge-Object Detection in Images)

VisDrone2018

说明：后续可能不更新了。

Tips: this repo will not be updated.

Baseline：

Name	maxDets	Result
Average Precision (AP) @( IoU=0.50:0.95)	maxDets=500	15.8738%.
Average Precision (AP) @( IoU=0.50 )	maxDets=500	21.7822%.
Average Precision (AP) @( IoU=0.75 )	maxDets=500	17.1753%.
Average Recall (AR) @( IoU=0.50:0.95)	maxDets= 1	0.83255%.
Average Recall (AR) @( IoU=0.50:0.95)	maxDets= 10	7.1636%.
Average Recall (AR) @( IoU=0.50:0.95)	maxDets=100	20.7602%.
Average Recall (AR) @( IoU=0.50:0.95)	maxDets=500	20.7602%.

Cases From Test DataSet:

Implementation of a method of data augmentation named patching:

before augmentation:

after augmentation(only patching pedestrian and awning-tricycle, show pedestrian only):

ECCV2018的一个workshop举办的比赛，详见Vision Meets Drones: A Challenge.

VisDrone2018_Dev_Kit: 官方提供的针对数据集的工具，用于评测。可以改为其他工具，比如在图片上显示BBox；

Txt2XML: 官方给定数据集的Ground Truth是自己的标注方式（Txt），该工具将该标注方式转化为PASCAL VOC2007的标注方式（XML）；Python实现；

ShowBBOXFromXML: 针对PASCAL VOC2007，在图片上显示BBox；Python实现；该工具已经和官方给定基于Matlab的代码做过准确度对比，检验通过；

数据集的Badcase: 将BBox画成一条线的，导致NaN错误，需要过滤，详情可以参考issues;

Code中分享了基于PyTorch的Faster R-CNN代码用于这个比赛，原始代码来自@jwyang，原始代码写的也有很多不完善的地方，但是是基于PyTorch实现的star最多的，用起来是没有问题的。Code中的仅仅作为该比赛代码的备份，不做正式分享。比如，没有数据。如果想要在现有代码基础上做些工作，可以联系我本人，帮助跑代码。

visdrone2018's Issues

你好，请问数据集能给一份吗？无人机的，[email protected]

pre trained model

hello,

Thank you for your contribution. My question is do you have some thing which is pre trained model.I can download that model and use that model in the code provided to detect vehicle from drone videos
That could be great help
Thank you.

关于数据增强用于maskrcnn

您好！我在做关于小目标检测问题，想请教您问题，如何联系您？

请问要加入数据训练，应该修改哪部分代码呢

如题，谢谢你的时间。

How to run faster-rcnn from your code

Hi,

Can you give steps on how to run the pytorch faster rcnn code for visdrone from your repo? I am benchmarking faster-rcnn from different repos. Please let me know how to run this from your repo
Thank you very much

Visdrone数据集格式问题

请问哪里有该数据集标签的格式介绍吗？比如您的代码里求xmax = line[0] + line[2] - 1; 难道标注时line[0]不是从0开始计算的吗？

The badcase annotations

Hi,

Could you please tell which annotations are causing nan problems? I tried two or three detectors, none of them are converging.

Patching Data Augmentation

Hi @zhpmatrix,

You mention patching data augmentation that you used and obtained great results. Can you provide a link to a paper or code in which this has been implemented?

Thanks and Regards,
Karthik

关于Faster R-CNN的改进部分

您好：
我VisDrone网站上下载了训练集，发现图片分辨率很大(1920,1080)，请问你们有做处理吗？还是直接丢进网络训练？