Results on the two outdoor and indoor sequences
(The videos are in xvid codec)
The results are definitely improved compare to the previous local model.
However, quadratic approximation may not be accurate, especially for the outdoor sequence with trace leaving behind when the foreground move away (the color trajectories below give an intuitive explanation for this inaccuracy).
| Indoor sequence | Previous model (local linear, rgb space): 3bandmodel\indoor\rgb_delta_tangent.avi | New model (global, quadratic, separate band):3bandmodel\indoor\3-band_global_quadratic.avi |
| Outdoor sequence | Previous model (local linear, rgb space): 3bandmodel\outdoor\rgb_tangent_local.avi | New model (global, quadratic, separate band):3bandmodel\outdoor\3-band_global_quadratic.avi |
Some color trajectories
r = f(idx)
idx = median(r + g + b)/3
|
Images |
R |
G |
B |
|
|
|
|
|
|
indoor |
|
|
|
outdoor
Videos for final
tracking results for the ETI-SEO project
(Indeo Codec)
ETI-VS2-AP-11-C4smoothen_final_tracks_ids.avi
Evaluation Result
ETI-VS2-AP-11-C7smoothen_final_tracks_ids.avi Evaluation
Result
(airport sequences)
ETI-VS2-BE-19-C1smoothen_final_tracks_ids.avi
Evaluation Result
ETI-VS2-BE-19-C4smoothen_final_tracks_ids.avi
Evaluation Result
(building entrances
with glass doors)
ETI-VS2-MO-1-C1smoothen_final_tracks_ids.avi
Evaluation Result
(subway sequence with
dim light)
ETI-VS2-RD-10-C4smoothen_final_tracks_ids.avi
Evaluation Result
(road- outdoor
sequence with large illumination change)
ETI-VS2-RD-6-C7smoothen_final_tracks_ids.avi
Evaluation
Result
(road outdoor sequence, a street)
(Xvid Codec)
ETI-VS2-AP-11-C4smoothen_final_tracks_ids_xvid.avi
ETI-VS2-AP-11-C7smoothen_final_tracks_ids_xvid.avi
ETI-VS2-BE-19-C1smoothen_final_tracks_ids_xvid.avi
ETI-VS2-BE-19-C4smoothen_final_tracks_ids_xvid.avi
ETI-VS2-MO-1-C1smoothen_final_tracks_ids_xvid.avi
ETI-VS2-RD-10-C4smoothen_final_tracks_ids_xvid.avi
ETI-VS2-RD-6-C7smoothen_final_tracks_ids_xvid.avi
Raw tracks for
(new codec: ETI-VS2-RD-10-C4\raw_tracks2_xvid.avi)
Box tracker (a short sample) ETI-VS2-RD-10-C4\temp3.avi
(new codec:ETI-VS2-RD-10-C4\temp3.avi )
New background subtraction result
newbgr\short-ETI-VS2-RD-10-C4-out_double_xvid.avi
VS2 (Data set 2 new
one)
|
ETI-VS2-AP-11-C4 (cam 4) |
||
|
ETI-VS2-AP-11-C7 (cam 7) |
|
ETI-VS2-BE-19-C1 (cam 1) |
|
|
|
ETI-VS2-BE-19-C4 (cam 4) |
|
|
ETI-VS2-MO-1-C1 |
|
|
ETI-VS2-RD-10-C4 |
Quick adaptation bgr_quick_update_ep1_25_alpha_0.7_beta_1.2.avi (produce very bad results because of the fast lighting changes) |
Slow adaptation |
|
ETI-VS2-RD-6-C7 |
Slow adaptation bgr_slow_update__ep1_30_alpha_0.8_beta_1.4.avi (Complete) |
1_390_ep1_45_alpha_0.8_beta_1.4.avi (incomplete only from frame 1 to frame 390, after that
the program crashes) |
VS1 (Data set 1)
The first two come from two cameras on the same scene, and can be assumed that empty scene is available.
The third one comes from a single camera and no empty scene is available.
For all sequence frames from 1-100 are used to build the background model.
|
ETI-VS1-AP-3-C4 |
||
|
ETI-VS1-AP-3-C7 |
||
|
ETI-VS1-MO-7-C1 |
||
|
Some observations |
Decrease alpha to less than .8 and increase beta more than 1.5 will make many part of the object missed, and obj become more and more fragmented. |
Increase alpha more than .85 and decrease beta less than 1.45 will make more background (shadow, reflection, etc.) included |
More notes:
(1) Alpha and Beta specify how much the program covers shadow and highlight. Alpha is for shadow and Beta for highlight. Usually, set alpha = 0.5, beta = 2.0. In order to include more shadow [highlight] to backgrounds, decrease [increase] Alpha [Beta], for example, Alpha = 0.4 [Beta = 2.1] (and vice versa). For monochrome imagery, these two parameters control the detection level.(from Kyungnam Kims instruction manual)
(2) Please download the XviD codec, e.g. below, if
you dont have it installed on your computer
http://download.divxmovies.com/XviD-1.1.0-30122005.exe