Comparison of DNN-based object detectors

Structure

• converters contains scripts to convert groundtruth of the traffic videos (text format) to the format of PASCAL VOC.

• readers contains scripts to read detections and groundtruth represented in text format.

• utilities contains scripts to estimate detection/tracking quality and to perform visual inspection:

  • average_precision.py to calculate average precision (AP) and draw precision-recall curve.
  • true_positive_rate.py to compute true positive rate (TPR).
  • false_detection_rate.py to calculate false detection rate (FDR).
  • false_positives_per_frame.py to compute number of false positives per frame/image.
  • play_bboxes.py to show groundtruth and detections simultaneously.
  • play_tracks.py to show constructed tracks.
  • auxiliary scripts required for AP, TPR and FDR computation.

• auxiliary/ssd-detector contains scripts to install and to execute SSD.

• vehicle-detector is a video-based vehicle detection system.

  • detector is a package containing implementation of detection methods.
  • tracker is a package containing implementation of tracking methods.
  • video-detector is a package containing implementation of video-based detection algorithms (provide detection and tracking). video_analyzer.py is a starting point.
  • tests is a set of learning tests.

References

1. Liu W., Anguelov D., Erhan D., Szegedy C., Reed S., Fu Ch.-Y., Berg A.C. SSD: Single Shot MultiBox Detector. 2016. [https://arxiv.org/abs/1512.02325].
2. Sources of SSD [https://github.com/weiliu89/caffe/tree/ssd].