SCUT FIR Pedestrian Dataset

A new benchmark dataset and baseline for on-road FIR pedestrian detection

Description

The SCUT FIR Pedestrian Datasets is a large far infrared pedestrian detection dataset. It consist of about 11 hours-long image sequences ($\sim 10^6$ frames) at a rate of 25 Hz by driving through diverse traffic scenarios at a speed less than 80 km/h. The image sequences were collected from 11 road sections under 4 kinds of scenes including downtown, suburbs, expressway and campus in Guangzhou, China. We annotated 211,011 frames for a total number of 477,907 bounding boxes around 7,659 unique pedestrians.

Statistic of training and testing set

more details see statistic.xlsx.

Training Set

S0-S10

Label	Frames with anno.	Bounding boxes	Occluded BB	Unique	Avg. frames per obj.
walk person	48857	103239	31630	1613	64.00
ride person	40519	71793	8697	807	88.96
squat person	1520	1608	324	22	73.09
people	26385	38358	9805	612	62.68
person?	13059	16820	6943	496	33.91
people?	6855	8479	2457	153	55.42
Summary	70517	240297	59856	3703	64.89

Testing Set

S11-S20

Label	Frames with anno.	Bounding boxes	Occluded BB	Unique	Avg. frames per obj.
walk person	43421	90526	26185	1523	59.44
ride person	43153	86201	8689	1017	84.76
squat person	8010	837	184	14	59.79
people	24098	33572	8903	647	51.89
person?	17208	22483	8759	642	35.02
people?	3399	3991	1604	97	41.14
Summary	73115	237610	54324	3940	60.31

Download

videos GoogleDrive BaiduYun

annotations GoogleDrive BaiduYun (please contact with xzhewei@gmail.com)

Tool

Github
Seq video format. Data Format is compatible with Caltech Pedestrian Dataset Format
datatool. Evaluation/labeling code for our dataset which is based on Caltech Dataset.
toolbox. The datatool depended tool which is based on Piotr's Matlab Toolbox.

Baseline

Thermal Faster R-CNN (TFRCN) in xzhewei/TFRCN released

Benchmark Results

Algorithm List

	Reasonable All	Overall
RPN+BF	8.28	25.19
TFRCN	9.98	32.32
RPN (RPN+BF)	12.07	32.94
Faster R-CNN-vanilla	19.75	52.00
RPN-vanilla	34.87	61.20
ACF-T+THOG	43.70	62.11

Ther mertic is log-average miss rate over the range of [$10^{-2}$, $10^0$].

ROC

Contact

Please contact Zhewei Xu [xzhewei at gmail.com] with questions.

lihengMo/SCUT_FIR_Pedestrian_Dataset