Author Affiliations
Department of Computer Science, North China Electric Power University, Baoding, Hebei 071000, Chinashow less
Fig. 1. Siamese network architecture
Fig. 2. Framework of proposed algorithm
Fig. 3. Multi-scale sampling construction pyramid
Fig. 4. Part of the experiment tracking sequences
Fig. 5. Comparison of tracking results of 10 algorithm. (a) Jumping (MB、FM); (b) CarScale (SV、OCC、FM、IPR、OPR); (c) Couple (SV、DEF、FM、BC、OPR); (d) Soccer (IV、SV、OCC、MB、FM、IPR、OPR、BC); (e) Ironman (IV、SV、OCC、MB、FM、IPR、OPR、OV、BC、LR); (f) Girl (SV、OCC、IPR、OPR); (g)Skiing (IV、SV、DEF、IPR、OPR); (h) Liquor (IV、SV、OCC、FM、IPR、OPR、OV、BC); (i) Woman (IV、SV、OCC、DEF、MB、FM、OPR)
Fig. 6. Precision and success rate of 50 sequences. (a) Precision; (b) success rate
Fig. 7. Precision and success rate on FM and SV attributes. (a) Precision on FM attribute; (b) success rate on FM attribute; (c) precision plot on SV attribute; (d) success rate on SV attribute
Fig. 8. Tracking precision and success rate on (a)(b) IV, (c)(d) OCC, (e)(f) LR, (g)(h) DEF, (i)(j) BC, and (k)(l) OV attributes
Layer | Support | Chan.map | Stride | For exemplar | For search | Chans |
---|
| | | | 127×127 | 255×127 | ×3 | Conv1 | 11×11 | 96×3 | 2 | 59×59 | 123×127 | ×96 | Pool1 | 3×3 | | 2 | 29×29 | 61×61 | ×96 | Conv2 | 5×5 | 256×48 | 1 | 25×25 | 57×57 | ×256 | Pool2 | 3×3 | | 2 | 12×12 | 28×28 | ×256 | Conv3 | 3×3 | 384×256 | 1 | 10×10 | 26×26 | ×192 | Conv4 | 3×3 | 384×192 | 1 | 8×8 | 24×24 | ×192 | Conv5 | 3×3 | 256×192 | 1 | 6×6 | 22×22 | ×168 |
|
Table 1. Architecture of CNN of proposed algorithm
Tracker | Proposed | CF2 | DCFNet | Siamfc | HDT |
---|
Platform | GPU | GPU | GPU | GPU | GPU | FPS /(frame·s-1) | 42.3 | 11 | 65.94 | 58 | 10 | Real | Y | N | Y | Y | N |
|
Table 2. Tracking speed comparison of proposed algorithm and other algorithms