PROST - Parallel Robust Online Simple Tracking

 
  sequence lemming, frame 108

Sequence lemming , frame 108

sequence liquor, frame 441

Sequence liquor , frame 441

 
Paper
 

Abstract

Tracking-by-Detection is an increasingly popular approach in order to tackle the visual tracking task. Existing adaptive methods suffer from the drifting problem, since they rely on self-updates of an on-line learning method. In contrast to previous work that tackled this problem via reformulating the self-learning to either semi-supervised or multiple-instance learning, we show that augmenting an on-line learning method with complementary tracking approaches can lead to better results. In particular, we use a simple template model as a non-adaptive element, a novel optical-flow-based meanshift tracker as highly adaptive element and an on-line random forest as adaptive appearance-based learner. We combine these three trackers in a simple cascade. All of our system parts are chosen in order to run on GPUs or similar multi-core systems, which allows for near real-time performance. We show the superiority of our system over current state-of-the-art online tracking methods in several experiments.

Download

[pdf]   [bibtex]

Sequences and Code
 

Download

prost_dataset.rar (242.7 MB)

Description

This file contains the four tracking sequences board, box, lemming and liquor together with ground truth as well as several tracking results. You can find the following files for each sequence:

imgs/00000.jpg - imgs/xxxxx.jpg     The sequence as numbered *.jpg images
sequence.wmv     The sequence as wmv video file
sequence_gt.txt     Ground truth positions of the tracking rectangle given as [x]   [y]   [width]   [height]
sequence_prost.txt     Tracking result of PROST
sequence_FragTrack.txt     Tracking result of FragTrack
sequence_MIL_TR001_c.txt     Tracking result of MilTrack, note that there are several different runs
sequence_GRAD.txt     Tracking result of GRAD

 

Additionally, there are two Matlab scripts in the /matlab folder:

evalExperiment.m     Reads tracking result files and evaluates their performance
createVideo.m     Creates video files of the tracking results

Videos
 

These videos show the tracking results of our method on the provided datasets.

 

board

box

 
 

lemming

liquor

 
Results
 

In this table we want to compare the results of different tracking methods on our dataset. If you would like to have your method stated here, just send us your paper showing your results on this database.

The columns state two evaluation scores for each sequence: 'pascal' and 'distance'. The 'distance' score gives the average euclidian distance between tracking rectangle and ground truth rectangle. The 'pascal' distance states the percentage of frames, where the object detection score used in the PASCAL challenge exceeds 0.5 (see the paper as well as our code for an example). This score can be interpreted directly as the percentage of frames, where a tracker works correctly.
The overall score states the arithmetic mean over all four pascal score results: In other words, PROST is capable to track 80.2 percent of the whole dataset correctly.

Overall

board

box

lemming

liquor

Method

pascal

pascal

distance

pascal

distance

pascal

distance

pascal

distance

PROST [1]

80.4

75.0

39.0

90.6

13.0

70.5

25.1

85.4

21.5

MILTrack [2]

49.2

67.9

51.2

24.5

104.6

83.6

14.9

20.6

165.1

FragTrack [3]

66.0

67.9

90.1

61.4

57.4

54.9

82.8

79.9

30.7

ORF [4]

27.3

10.0

154.5

28.3

145.4

17.2

166.3

53.6

67.3

GRAD [5]

88.9

94.3

14.7

91.8

13.2

78.0

28.4

91.4

11.9

 

[1] Santner, Leistner, Saffari, Pock, Bischof: PROST: Parallel Robust Online Simple Tracking, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010, to appear, [www]

[2] Babenko, Yang, Belongie: Visual Tracking with Online Multiple Instance Learning, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009, [www]

[3] Adam, Rivlin, Shimshoni: Robust Fragments-based Tracking using the Integral Histogram, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2006, [www]

[4] Saffari, Leistner, Santner, Godec, Bischof: On-line Random Forests, 3rd IEEE ICCV Workshop on Online Computer Vision, 2009, [www]

[5] Klein, Cremers: Boosting Scalable Gradient Features for Adaptive Real-Time Tracking, Int. Conf. on Robotics and Automation (ICRA), 2011, [youtube]