Regionlets Generic Object Detection

Abstract

Regionlets is an object detection approach aimed at detecting objects in arbitrary scales, arbitrary viewpoints, with the capability of deformation and sub-category handling. It detects objects at their original scale, i.e. the approach does not resize the image when performing training/testing. Compared to traditional object detection paradigm, Regionlets model differs in:
  • Feature histograms are built in variable regions (vs fixed size cells, 8x8 HOG for example)
  • Feature extraction regions are normalized to detection windows.
  • Deformation handling is learned from data.
  • Learned Regionlets model is lot limited by a fixed scale or aspect ratio.
  • Figure 1 shows the detection framework.

    Post thumbnail
    Figure 1. Regionlets detection framework
    Figure 2 shows the difference of a normalized region and a traditional region in two examples(the original window and a tranformed one with double resolution).
    Post thumbnail
    Figure 2. Normalized feature extraction region
    Table 1 shows the detection performance on the PASCAL VOC 2007 dataset.
    Post thumbnail
    Table 1. Performance of regionlets on the PASCAL VOC 2007 dataset.

    Code

    Currently not available

    Demo

    Car detection and recognition

    Resources:

  • PAMI 2014 paper: Regionlets for Generic Object Detection
  • ICCV 2013 paper: Regionlets for Generic Object Detection
  • ACCV 2014 paper: Accurate Object Detection with Location Relaxation and Regionlets Relocalization
  • BMVC 2014 paper: Generic Object Detection with Dense Neural Patterns and Regionlets
  • Detection results on KITTI dataset: Tracking DetectionDetection DownloadTracking Download
  • Technic Report: Generic Object Detection with Dense Neural Patterns and Regionlets
  • Oral presentation
  • ImageNet(ILSVRC2013) presentation
  • Video demonstration
  • Longer version of the paper will be available soon.