无人机视觉 drone

http://www.aiskyeye.com/

2018年已经办过一年了。2019年在ICCV上办。

We encourage the participants to use the provided training data for each task, but also allow them to use additional training data. The use of additional training data must be indicated in the "method description" when uploading results to the server.

We emphasize that any form of annotation or use of the VisDrone testing sets for either supervised or unsupervised training is strictly forbidden. The participants are required to explicitly specify any an all external data used for training in the "method description" in submission. In addition, the participants are NOT allowed to train a model in one task using the training or validation sets in other tasks.

we have divided the test set into two splits, including test-challenge and test-dev. Test-dev一天交3次，test-challenge一共交3次。

不明白为什么download里test-dev是不能下载的。。。

 

Vision Meets Drones: A Challenge

Abstract

our benchmark has more than 2:5 million annotated instances in 179; 264 images/video frames.

Introduction

现状：缺少大的数据集。

Altogether we carefully annotated more than 2:5 million bounding boxes of object instances from these categories. Moreover, some important attributes including visibility of scenes, object category and occlusion, are provided for better data usage.

感觉还是挺大的，也没什么别的数据集可以利用了。

3.2 Task 1: Object Detection in Images

The VisDrone2018 provides a dataset of 10; 209 images for this task, with 6; 471 images used for training, 548 for validation and 3; 190 for testing.

For truncation ratio, it is used to indicate the degree of object parts appears outside a frame. If an object is not fully captured within a frame, we annotate the bounding box across the frame boundary and estimate the truncation ratio based on the region outside the image. It is worth mentioning that a target is skipped during evaluation if its truncation ratio is larger than 50%.

task1 遮挡情况

Three degrees of occlusions: no occlusion (occlusion ratio 0%), partial occlusion (occlusion ratio 1% 50%), and heavy occlusion (occlusion ratio > 50%).

2018task1winner

选择retina net，移除后面两层，只用P3,4,5

数据分布

 

Cascade R-CNN: Delving into High Quality Object Detection

Single-Shot Bidirectional Pyramid Networks for High-Quality Object Detection

数据观察

task1

bbox观察

数据分布

bbox尺度分布

	<p2	p2	p3	p4	p5	p6	>p6
train	48.3%	30.6%	16.2%	4.4%	0.5%	0.01%	0%
val	55.3%	30.8%	11.4%	2.4%	0.1%	0.007%	0%
train_val	49.0%	30.6%	15.7%	4.1%	0.5%	0.01%	0%

 

bbox尺度*1.5分布

	<p2	p2	p3	p4	p5	p6	>p6
train	28.2%	34.6%	24.5%	10.5%	2.0%	0.1%	0.002%
val	33.0%	37.6%	21.8%	6.9%	0.7%	0.03%	0.005%
train_val	28.7%	34.9%	24.2%	10.2%	1.9%	0.1%	0.002%

bbox尺度*3分布

	<p2	p2	p3	p4	p5	p6	>p6
train	6.2%	22.0%	34.6%	24.5%	10.6%	2.0%	0.1%
val	7.1%	26.0%	37.6%	21.8%	6.9%	0.7%	0.03%
train_val	6.3%	22.4%	34.9%	24.2%	10.2%	1.9%	0.1%

类别分布

类别数量分布

	0 ignored	1 pedestrain	2 people	3 bicycle	4 car	5 van	6 truck	7 tricycly	8 awning-tricycle	9 bus	10 motor	11 others
train	2.49%	22.44%	7.65%	2.96%	40.97%	7.06%	3.64%	1.36%	0.92%	1.68%	8.39%	0.43%
val	3.43%	22.02%	12.76%	3.24%	35.01%	4.92%	1.87%	2.60%	1.32%	0.62%	12.16%	0.08%
trian_val	2.59%	22.40%	8.17%	2.99%	40.37%	6.84%	3.46%	1.49%	0.96%	1.57%	8.77%	0.40%

 

类别面积分布

	0 ignored	1 pedestrain	2 people	3 bicycle	4 car	5 van	6 truck	7 tricycly	8 awning-tricycle	9 bus	10 motor	11 others
train	3.89%	6.42%	1.82%	1.45%	54.11%	11.66%	9.46%	1.39%	1.07%	4.69%	3.35%	0.69%
val	8.23%	7.44%	3.67%	1.42%	54.88%	7.68%	5.80%	2.38%	1.29%	1.82%	5.23%	0.15%
trian_val	4.18%	6.49%	1.94%	1.45%	54.16%	11.39%	9.22%	1.46%	1.09%	4.50%	3.47%	0.65%

图片尺寸分布

	480×360	960×540	1344×756	1360×765	1389×1042	1398×1048	1400×788	1400×1050	1916×1078	1920×1080	2000×1500	all
train	1	250	1	743	1	30	1299	2498	537	339	772	6462
val	0	121	0	408	0	0	0	0	0	19	0	548
test	0	64	0	150	0	0	712	491	127	36	0	1580

train 平均图像大小：1575738.0529247911

 

b0说task2里同一个物体的类别会变。

Pyramid Scene Parsing Network