Skip to content

Experimenting with Person Re-ID with deep learning libraries

Notifications You must be signed in to change notification settings

BonaventureR/person-reid

Repository files navigation

Person Re-ID

This repo experiments with cross-data evaluation from pre-existing datasets such as Market1501 and CUKH-03 on our collected dataset.

These experiments utilize the TorchReid Library created by Kaiyang Zhou https://kaiyangzhou.github.io/deep-person-reid/ as a means to test Cross-Entropy and Triplet Loss Function's effects on mAP, and rank accuracies on collected datasets.

Training Data

alt text

Market-1501 dataset annotates 27 attributes, containing 751 identities for training and 750 for testing, that are annotated in the identity level. Thus, the file contains 27 x 751 attributes for training and 27 x 750 for test.

These images are used for cross-data evaluation on our dataset containing two disjoint cameras in various variable environment settings (lighting, view, background).

Ranking@10 Image Preview

Softmax + CE Loss (Random Crop, Random Flip)

alt text

Softmax + CE Loss (Random Crop, Random Flip, Random Erase, Color Jitter)

alt text

Triplet Loss

alt text

But also, performed poorly in some examples,

alt text

alt text

Evaluation

Softmax+CLE (Crop, Flip) Softmax+CLE(Crop, Flip, CJitter, Patch) TripletLoss (Crop, Flip)
mAP 45.6% 92.9% 95.4%
Rank@1 10.1% 97.1% 95.6%
Rank@5 25.3% 100.0% 98.5%
Rank@10 53.2% 100.0% 98.5%
Rank@20 63.3% 100.0% 98.5%

References

[1] Hermans, A., Beyer, L., & Leibe, B. (2017). In Defense of Triplet Loss for Person Re-Identification. Retrieved December 15, 2020, from https://arxiv.org/pdf/1703.07737.pdf

[2] Li, W., Zhao, R., Xiao, T., & Wang, X. (2014). DeepReID: Deep Filter Pairing Neural Network for Person Re-identification. 2014 IEEE Conference on Computer Vision and Pattern Recognition. doi:10.1109/cvpr.2014.27

[3] Wang, G., Lai, J., Huang, P., & Xie, X. (2018). Spatial-Temporal Person Re-identification. Retrieved December 15, 2020, from https://arxiv.org/pdf/1812.03282v1.pdf

[4] Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., & Tian, Q. (2015). Scalable Person Re-Identification: A Benchmark. Retrieved December 16, 2020, from https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Zheng_Scalable_Person_Re-Identification_ICCV_2015_paper.pdf

[5] Zhong, Z., Zheng, L., Zheng, Z., Li, S., & Yang, Y. (2018). Camera Style Adaptation for Person Re-identification. Retrieved December 15, 2020, from https://arxiv.org/pdf/1711.10295.pdf

[6] Li, W., Zhao, R., Xiao, T., & Wang, X. (2014). DeepReID: Deep Filter Pairing Neural Network for Person Re-identification. 2014 IEEE Conference on Computer Vision and Pattern Recognition. doi:10.1109/cvpr.2014.27

[7] Hirzer, M., Beleznai, C., Roth, P., & Bischof, H. (2011). Person Re-identification by Descriptive and Discriminative Classification. SCIA.

About

Experimenting with Person Re-ID with deep learning libraries

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published