EuroCity Persons Dataset

Back

Submission #35, by Qihua Cheng

Used Method

FasterRCNN with Multi-Stream RCNN

Submitted on

2019-07-08 08:48:13

Description

We use multi-stream RCNNs to handle the proposals given by an RPN. We use multi-scale training at training time. At test time, only one single scale is used to test. We filter samples that are more than 40% occluded or smaller than 20 pixels in height. We adopt the same setting with Adapted FasterRCNN. We use VGG16 as our based network. We use pedestrian and rider to train.

Method description: We use multi-stream RCNNs to handle the proposals given by an RPN. We use multi-scale training at training time. At test time, only one single scale is used to test. We filter samples that are more than 40% occluded or smaller than 20 pixels in height. We adopt the same setting with Adapted FasterRCNN. We use VGG16 as our based network. We use pedestrian and rider to train.
Up-/Downsampling: 1.0
Computation Environment: Intel Xeon E5-260, Nvidia TitanX
Computation Time: 0.36

Result

night
0.150
0.253
0.653
0.295
0.135
0.237
0.629
0.277