SPID: Surveillance Pedestrian Image Dataset


As a subset of BEST2016 dataset, this dataset was collected for the pedestrian detection research works using surveillance images or videos. The images in Surveillance Pedestrian Image Dataset are extracted from diverse set of videos recorded by on-using surveillance cameras. They have been selected to cover a wide range of pedestrian detection challenges and are representative of typical outdoor surveillance scenarios. SPID totally consists of 14550 training images and 15439 test images, comprising a total of 29989 original images and 110069 labeled pedestrians. All images are color and saved as .jpg. The dataset is consisted of two formats: (a) original images with corresponding annotation files, and (b) individual pedestrian images in various scales. SPID includes the following pedestrian challenges: diverse scenes, dynamic background, different poses and views, various illumination, small scales. The research is described in detail in ACCV2016 paper “SPID: Surveillance Pedestrian Image Dataset and Perfomance Evaluation for Pedestrain Detection”.


Fig. 1: Examples of diverse pedestrians in SPID


Ground Truth

To enable a precise quantitative comparison and ranking of various algorithms, all of our images come with accurate ground-truth and annotation of pedestrian coordinates. Each pedestrian is described with one bounding box (Xmin, Ymin, Xmax, Ymax), see label <bndbox> in xml files. If one image contains multiple pedestrians, each pedestrian is assigned one ID label independently.



The training data consists of six sets (set00-set05), and the first five sets are ~1GB size each, and each set contains both the original images and their annotation files. Set05 contains 9150 individual pedestrian images. The testing data consists of five image sets (set06-set10, ~1GB each), each set has about 3000 original images, along with annotation files.

Download (SPID2016V2):

Training Data

Testing Data

1. SPID2016V1
  • Release date: 2016/07/27
  • 14550 training images and 15439 test images
  • 29989 original images and 110069 labeled pedestrians.

2. SPID2016V2
  • Change date: 2016/12/30
  • 14420 training images and 15277 test images
  • 29697 original images and 110018 labeled pedestrians
  • Change detail:

1). Remove 162 images with no pedestrian object from set07, 130 images from set01.

2). Supplement some missing pedestrian objects in both testing & training sets, ~400 xml are updated.

We would like to thank the following individuals for their contributions to make this dataset more accurate:
  • Yi Xu, Shanghai Jiao Tong University, Shanghai China
  • Junjie Bai, Qual Comm. Co., America


Usage Policy

  • No commercial reproduction, distribution, display or performance rights in this work are provided.
  • If you use this facility in any publication, we request you to kindly acknowledge this website ( and cite the following paper:
    Dan Wang, Chongyang Zhang, Hao Cheng, Yanfeng Shang, and Lin Mei, “SPID: Surveillance Pedestrian Image Dataset and Perfomance Evaluation for Pedestrain Detection”, in the 13th Asian Conference on Computer Vision Workshop on Benchmark and Evaluation of Surveillance Task (ACCV BEST2016), Taipei ROC, November 20-24, 2016. (pdf)

2016 © SJTU-BEST 沪交ICP备20160083

Support by : Wei Cheng