Jingjing Liu (Danny)
Department of Computer Science, Rutgers University,
110 Frelinghuysen Road, Piscataway, NJ 08854-8019
Email: jl1322 at cs dot rutgers dot edu
[Home] [Education] [Working Experience] [Projects] [Publications] [TA Work] [Courses] [Others]


Multispectral Deep Neural Networks for Pedestrian Detection
Figure.1. Top: the original annotations. Bottom: our improved annotations.
Description
Some problematic annotations were found in KAIST test set, e.g., missing annotation of valid human instances and incorrect labeling on background objects. Obviously, such mistakes in ground truth would lead to error-prone evaluations of pedestrian detectors. Hence, it would be unreliable to make any conclusions on these detectors, based on such annotations. To obtain reliable ground truth, we manually labelled every image in KAIST test set. In the following figure, we show the statistics of the improved annotations on 2,252 images in KAIST test set.
Figure.2. Statistics of the improved annotations on 2,252 images in KAIST test set.

Download
  • KAIST Multispectral Pedestrian Dataset. (Link to KAIST dataset)
  • Improved annotations of test set of KAIST. (Click to download)
  • Results of different models tested on test set of KAIST. (Click to download)
  • Code
  • Link to project on GitHub.(https://github.com/denny1108/multispectral-pedestrian-py-faster-rcnn)

  • References
  • Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N. Metaxas. Exploiting Deep Neural Network Fusion for Robust Multispectral Pedestrian Detection. (Journal version of BMVC'16 paper, in submission).
  • Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N. Metaxas. Multispectral Deep Neural Networks for Pedestrian Detection. BMVC, 2016.
  • Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi and In So Kweon. Multispectral Pedestrian Detection: Benchmark Dataset and Baseline. CVPR, 2015.