Discriminatively Trained Deformable Part Models with Unsupvervised Training [LSVM-MDPM-us]

Submitted on 25 Oct. 2012 14:09 by
Philip Lenz (KIT)

Running time:10 s
Environment:4 cores @ 3.0 Ghz (C/C++)

Method Description:
Version 4 of the popular object detector (16 classes for cars, 8 classes for pedestrians, 4 classes for cyclists). Additionally, a large-enough margin around the horizon has been enlarged by factor 3 to also detect smaller objects.
Negative training data is harvested from the positive images. Negative bounding boxes are only used, if there overlap with a positive bounding box is less than 20%.
As training set we have used all available fully visible or partly occluded bounding boxes per orientation. Otherwise, the default parameters of the algorithm have been used.
Latex Bibtex:
author = {Pedro F. Felzenszwalb and Ross B. Girshick and David McAllester and Deva Ramanan},
title = {Object Detection with Discriminatively Trained Part-Based Models},
journal ={PAMI},
volume = {32},
year = {2010},
pages = {1627-1645}

Detailed Results

Object detection and orientation estimation results. Results for object detection are given in terms of average precision (AP) and results for joint object detection and orientation estimation are provided in terms of average orientation similarity (AOS).

Benchmark Easy Moderate Hard
Car (Detection) 70.52 % 56.10 % 42.87 %
Pedestrian (Detection) 50.01 % 38.35 % 34.78 %
Cyclist (Detection) 40.31 % 30.81 % 28.17 %
This table as LaTeX

This figure as: png eps pdf txt gnuplot

This figure as: png eps pdf txt gnuplot

This figure as: png eps pdf txt gnuplot

eXTReMe Tracker