456
Z. Zhang et al.
Batch size. We test different batch sizes from 8 to 64, and Fig. 1 shows the
detailed results. From Fig. 1, we can see that it is suitable for the baseline to set
32 batch sizes and obtain the best results. It is harmful for the performance of
the baseline when setting small batch size.
Basic learning rate. We initialize the basic learning rate with different values,
such as 0.001, 0.005, 0.01 and so on. Figure 2 shows experimental results and
we could see that the baseline may learn worse feature representations when the
basic learning rate is initialized with large value or small value.
4 Conclusion
In this paper, we have evaluated the performance of person Re-ID baseline under
different implementation details. We modify ResNet-50 as the person Re-ID baseline by changing original FC layer. We have shown that these implementation
details could have influence on the performance of the baseline. Our work is
helpful and meaningful for training deep model in the person Re-ID community.
Acknowledgements. This work was supported by National Natural Science Foundation of China under Grant Nos. 61501327 and 61711530240, Natural Science Foundation
of Tianjin under Grant No. 17JCZDJC30600, the Fund of Tianjin Normal University
under Grant No. 135202RC1703, the Open Projects Program of National Laboratory
of Pattern Recognition under Grant No. 201800002, and the Tianjin Higher Education
Creative Team Funds Program.
References
1. Wang X (2013) Intelligent multi-camera video surveillance: a review. Pattern
Recognit Lett 34(1):3–19
2. Loy CC, Xiang T, Gong S (2009) Multi-camera activity correlation analysis. In:
IEEE conference on computer vision and pattern recognition, Miami, pp 1988–1995
3. Zhang Z, Wang C, Xiao B, Zhou W, Liu S (2012) Action recognition using contextconstrained linear coding. IEEE Signal Process Lett 19(7):439–442
4. Zhang Z, Wang C, Xiao B, Zhou W, Liu S, Shi C (2013) Cross-view action recognition via a continuous virtual path. In: IEEE conference on computer vision and
pattern recognition, Portland, pp 2690–2697
5. Gao S, Wang C, Xiao B, Shi C, Zhang Z (2014) Stroke bank: a high-level representation for scene character recognition. In: International conference on pattern
recognition, Stockholm, pp 2909–2913
6. Zhang Z, Mei X, Xiao B (2015) Abnormal event detection via compact low-rank
sparse learning. IEEE Intell Syst 31(2):29–36
7. Gheissari N, Sebastian TB, Hartley R (2006) Person reidentification using spatiotemporal appearance. In: IEEE conference on computer vision and pattern
recognition, vol 2, New York, pp 1528–1535
8. Farenzena M, Bazzani L, Perina A, Murino V, Cristani M (2010) Person reidentification by symmetry-driven accumulation of local features. In: IEEE conference on computer vision and pattern recognition, San Francisco, pp 2360–2367
Z. Zhang et al.
Batch size. We test different batch sizes from 8 to 64, and Fig. 1 shows the
detailed results. From Fig. 1, we can see that it is suitable for the baseline to set
32 batch sizes and obtain the best results. It is harmful for the performance of
the baseline when setting small batch size.
Basic learning rate. We initialize the basic learning rate with different values,
such as 0.001, 0.005, 0.01 and so on. Figure 2 shows experimental results and
we could see that the baseline may learn worse feature representations when the
basic learning rate is initialized with large value or small value.
4 Conclusion
In this paper, we have evaluated the performance of person Re-ID baseline under
different implementation details. We modify ResNet-50 as the person Re-ID baseline by changing original FC layer. We have shown that these implementation
details could have influence on the performance of the baseline. Our work is
helpful and meaningful for training deep model in the person Re-ID community.
Acknowledgements. This work was supported by National Natural Science Foundation of China under Grant Nos. 61501327 and 61711530240, Natural Science Foundation
of Tianjin under Grant No. 17JCZDJC30600, the Fund of Tianjin Normal University
under Grant No. 135202RC1703, the Open Projects Program of National Laboratory
of Pattern Recognition under Grant No. 201800002, and the Tianjin Higher Education
Creative Team Funds Program.
References
1. Wang X (2013) Intelligent multi-camera video surveillance: a review. Pattern
Recognit Lett 34(1):3–19
2. Loy CC, Xiang T, Gong S (2009) Multi-camera activity correlation analysis. In:
IEEE conference on computer vision and pattern recognition, Miami, pp 1988–1995
3. Zhang Z, Wang C, Xiao B, Zhou W, Liu S (2012) Action recognition using contextconstrained linear coding. IEEE Signal Process Lett 19(7):439–442
4. Zhang Z, Wang C, Xiao B, Zhou W, Liu S, Shi C (2013) Cross-view action recognition via a continuous virtual path. In: IEEE conference on computer vision and
pattern recognition, Portland, pp 2690–2697
5. Gao S, Wang C, Xiao B, Shi C, Zhang Z (2014) Stroke bank: a high-level representation for scene character recognition. In: International conference on pattern
recognition, Stockholm, pp 2909–2913
6. Zhang Z, Mei X, Xiao B (2015) Abnormal event detection via compact low-rank
sparse learning. IEEE Intell Syst 31(2):29–36
7. Gheissari N, Sebastian TB, Hartley R (2006) Person reidentification using spatiotemporal appearance. In: IEEE conference on computer vision and pattern
recognition, vol 2, New York, pp 1528–1535
8. Farenzena M, Bazzani L, Perina A, Murino V, Cristani M (2010) Person reidentification by symmetry-driven accumulation of local features. In: IEEE conference on computer vision and pattern recognition, San Francisco, pp 2360–2367