People Detection benchmark repository
Results
Author: Navneet Dalal.
Contact university/company: INRIA Rhone-Alps, Montbonnot, France
Contact email : Navneet.Dalal@inrialpes.fr
Method name: Histograms of Oriented Gradients (HOG) .
Description and pocessing time: The HOG detector is based on exhaustive search and a holistic person model using the Histogram of Oriented Gradients. It consists in scanning the full image looking for similarities with the chosen person model, evaluating different detection windows with a classifier at multiple scales and locations. The HOG approach is implemented in C++ and the computational cost is around 1 second per frame with 352x288 images (there is a faster implementation in OpenCV that runs around 0.1 seconds per frame). The tests have been performed on a Pentium IV with a CPU frequency of 2.4 GHz and 3GB RAM.
Parameters: Author default parameters here
You can download the people detection results of this method by clicking here.
Results for each video sequence:
Video |
HOG |
1 |
89.3 |
2 |
63.2 |
3 |
55.6 |
4 |
10.1 |
5 |
61.3 |
6 |
49.9 |
7 |
0.0 |
8 |
0.0 |
9 |
47.4 |
10 |
7.2 |
11 |
10.7 |
12 |
82.0 |
13 |
70.9 |
14 |
13.6 |
15 |
46.5 |
16 |
21.1 |
Average AUC |
39.3 |
Ranking |
5.19 |
Results for each background complexity:
Background complexity |
HOG |
Baseline |
69.4 |
Dynamic Background |
35.7 |
Camera Jitter |
25.0 |
Intermittent Object Motion |
16.3 |
Shadow |
46.8 |
Average AUC |
38.6 |
Ranking |
5.60 |
Results for each classification complexity:
Classification complexity |
HOG |
Low |
62.3 |
Medium |
53.3 |
High |
6.9 |
Average AUC |
40.8 |
Ranking |
5.67 |
