PICASO: PIxel Correspondences And SOft Match Selection for Fast Tracking

Radu Timofte, Junseok Kwon, Luc Van Gool
Journal Computer Vision and Image Understanding (CVIU 2016), 2016

This is the supplementary material with results (videos, recall and CLE values) for the PICASO tracker proposed in our CVIU 2016 paper.
You can download the archived supplementary material from here.

Tracking Speed The processing speed is important in practice. In Table 3 we directly computed the speed for MTT and our PICASO, while for the other methods we report the values from other independent studies. The FoT speed is from (Vojir and Matas, 2014), that for MS f g from (Vojir et al., 2013), for ALIEN from (Pernici and Del Bimbo, 2014), for PixelTrack from (Duffner and Garcia, 2013), and the speeds for Struck, HT and TLD are taken from (Kristan et al., 2013). PICASO, FoT, MS, PixelTrack can reach more than 30 fps on desktop CPUs. However, our PICASO is Matlab implemented, while the other methods use C++. We expect that a C++ implementation of PICASO can be several times faster than the current unoptimized Matlab implementation. Reading the images and extracting the pixel templates are time bottlenecks of our PICASO tracker, in comparison INN matching takes less than half. For exemple, the highest INN matching time was for `Panda', with a model of 230 pixels, an average of 25 fps, and CLE=3.5: reading the images takes 16.97% of the time, 600 pixels extraction -48.04%, while matching -29.18%.

The settings are the ones from the paper.

We include also:
- results obtained with and without occlusion masks, showing the benefit of them.
- results for different block sizes for the pixel templates.
- results for different sizes for the normalized object model window.


(Since it is a re-run there might be some slight differences compared to the paper.)

Please click on the sequence titles to see the videos with the result of PICASO.
We worked with 28 video sequences for a total of 10992 processed frames.

The "search window" from the paper is marked as "gate" in the videos.




Babenko videos: PICASO tracker recall and average center location error (CLE). A total of 5017 frames. See Table 1, in the paper.

Recall (>0.1)Recall (>0.5)CLE
David100.00100.005.63
Sylvester100.0087.318.59
Girl100.0088.0018.74
Face Occ. 1100.00100.004.68
Face Occ. 2100.0087.6515.67
Coke100.0096.554.33
Tiger 194.2972.8610.54
Tiger 291.6756.9412.39
average98.2586.1610.07

Godec videos: PICASO tracker recall and average center location error (CLE). A total of 2385 frames. See Table 2, in the paper.

Recall (>0.1)Recall (>0.5)CLE
Cliff-dive 1100.0065.3317.64
Motocross 1100.0058.2812.41
Skiing15.007.50215.87
Mountain-bike100.00100.0011.19
Cliff-dive 2100.0015.0015.21
Volleyball100.0054.715.97
Motocross 2100.00100.008.18
Transformer100.0044.7261.01
Diving86.1824.8829.04
High Jump100.0029.0910.07
Gymnastics100.0091.6411.27
average91.0253.7436.17

Training videos: PICASO tracker recall and average center location error (CLE). The results obtained without using occlusion masks are in the brackets. INN (lambda=0.25), maximum 600 pixels, PICASO settings from the paper. A total of 3590 frames.

Recall (>0.1)Recall (>0.5)CLE
Panda100.00 ( 68.30 ) 86.99 ( 50.28 ) 4.06 ( 39.06 )
Surfer 98.67 ( 98.67 ) 80.00 ( 92.00 ) 6.45 ( 5.27 )
Coupon Book100.00 ( 100.00 )100.00 ( 100.00 )4.99 ( 4.88 )
Car11100.00 ( 91.03 ) 83.33 ( 84.62 ) 4.80 ( 6.04 )
Jumping 78.21 ( 11.54 ) 59.94 ( 9.94 ) 15.58 ( 59.42 )
Deer100.00 ( 100.00 ) 88.57 ( 88.57 ) 9.84 ( 9.50 )
Singer 99.14 ( 99.43 ) 30.86 ( 30.86 ) 11.51 ( 11.15 )
Torus 44.49 ( 38.78 ) 16.73 ( 17.11 ) 56.46 ( 61.20 )
Woman 97.15 ( 99.83 ) 91.61 ( 93.79 ) 9.88 ( 3.90 )
average90.85 ( 78.62 )70.89 ( 63.02 )13.73 ( 22.27 )

Training videos: PICASO tracker recall and average center location error (CLE) for 3x3/5x5/7x7/9x9 block size pixel templates, respectively. INN (lambda=0.25), maximum 600 pixels. A total of 3590 frames. The best average performance is achieved for 5x5 blocks, as used in the paper.

Recall (>0.1)Recall (>0.5)CLE
Panda 69.08 /100.00 / 99.89 / 68.63 57.51 / 86.99 / 80.65 / 49.94 39.55 / 4.06 / 4.77 / 36.79
Surfer 97.33 / 98.67 /100.00 /100.00 88.00 / 80.00 / 93.33 / 89.33 6.38 / 6.45 / 4.19 / 4.97
Coupon Book 100.00 /100.00 /100.00 /100.00 100.00 /100.00 /100.00 /100.00 4.79 / 4.99 / 4.88 / 5.51
Car11 96.15 /100.00 /100.00 /100.00 80.77 / 83.33 /100.00 /100.00 5.83 / 4.80 / 2.07 / 1.91
Jumping 11.86 / 78.21 / 13.78 / 16.03 8.65 / 59.94 / 11.22 / 13.46 91.96 / 15.58 / 44.73 / 48.73
Deer 100.00 /100.00 /100.00 /100.00 88.57 / 88.57 / 88.57 / 88.57 8.38 / 9.84 / 9.53 / 9.07
Singer 78.86 / 99.14 / 99.43 / 99.43 23.14 / 30.86 / 30.86 / 30.86 28.39 / 11.51 / 10.46 / 9.41
Torus 44.49 / 44.49 / 39.54 / 42.21 17.87 / 16.73 / 12.55 / 9.13 54.36 / 56.46 / 63.21 / 58.08
Woman 21.48 / 97.15 / 98.32 / 98.32 18.46 / 91.61 / 93.62 / 93.29 137.06 / 9.88 / 6.55 / 6.94
average 68.80 / 90.85 / 83.44 / 80.51 53.66 / 70.89 / 67.87 / 63.84 41.85 / 13.73 / 16.71 / 20.16

Training videos: PICASO tracker recall and average center location error (CLE) for 256/512/1024/2048/4096 normalized object model window size, respectively, when the initialization object window size exceeds twice this size. INN (lambda=0.25), maximum 600 pixels, 5x5 block size. A total of 3590 frames. The best average performance is achieved for 1024 and 2048. For PICASO we used 1024 in the paper.

Recall (>0.1)Recall (>0.5)CLE
Panda 99.67 /100.00 /100.00 /100.00 /100.00 48.83 / 86.99 / 86.99 / 86.99 / 86.99 6.43 / 4.06 / 4.06 / 4.06 / 4.06
Surfer 98.67 / 97.33 / 98.67 / 98.67 / 98.67 86.67 / 78.67 / 80.00 / 80.00 / 80.00 6.43 / 7.92 / 6.45 / 6.45 / 6.45
Coupon Book 100.00 /100.00 /100.00 /100.00 / 40.00 100.00 /100.00 /100.00 / 98.46 / 38.46 8.14 / 6.43 / 4.99 / 6.19 / 66.71
Car11 66.67 /100.00 /100.00 /100.00 /100.00 64.10 / 83.33 / 83.33 / 83.33 / 83.33 24.79 / 4.80 / 4.80 / 4.80 / 4.80
Jumping 16.99 / 31.09 / 78.21 / 78.21 / 78.21 12.82 / 16.99 / 59.94 / 59.94 / 59.94 114.17 / 32.76 / 15.58 / 15.58 / 15.58
Deer 100.00 /100.00 /100.00 /100.00 / 32.86 90.00 / 88.57 / 88.57 / 85.71 / 22.86 9.59 / 10.68 / 9.84 / 10.41 /130.43
Singer 100.00 / 99.71 / 99.14 / 96.29 / 75.14 32.86 / 30.57 / 30.86 / 30.00 / 29.43 9.15 / 11.17 / 11.51 / 12.48 / 35.63
Torus 40.30 / 33.46 / 44.49 / 48.29 / 48.29 19.39 / 7.98 / 16.73 / 27.38 / 27.38 60.01 / 67.01 / 56.46 / 46.51 / 46.51
Woman 38.59 / 21.98 / 97.15 / 95.97 / 95.97 20.97 / 20.81 / 91.61 / 81.04 / 81.04 124.91 /139.34 / 9.88 / 13.41 / 13.41
average 73.43 / 75.95 / 90.85 / 90.82 / 74.35 52.85 / 57.10 / 70.89 / 70.32 / 56.60 40.40 / 31.57 / 13.73 / 13.32 / 35.95



11-Jul-2015 13:41