pith. sign in

arxiv: 1907.03253 · v1 · pith:44RS32QQnew · submitted 2019-07-07 · 💻 cs.CV

A Novel Teacher-Student Learning Framework For Occluded Person Re-Identification

Jiaxuan Zhuo , Jianhuang Lai , Peijia Chen This is my paper

Pith reviewed 2026-05-25 01:32 UTC · model grok-4.3

classification 💻 cs.CV
keywords person re-identificationocclusionteacher-student learningco-saliencydomain adaptationdeep learningcomputer vision
0
0 comments X

The pith

A teacher-student framework learns occlusion-robust person re-identification by transferring from full-body images using simulated occlusions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops a teacher-student learning framework to create models that identify people even when parts of their bodies are blocked from view. The teacher network starts with abundant full-body person images and uses a simulator to add more occlusions gradually while a co-saliency network highlights important regions automatically. The student network then adapts this knowledge using the limited available real occluded images. A sympathetic reader would care because person re-identification is widely used in surveillance and the method reduces the need for large occluded training sets.

Core claim

The central claim is that equipping a teacher network with a co-saliency network for feature extraction and identity classification plus a cross-domain simulator for generating occlusions on full-body data allows effective knowledge transfer to a student network trained on real occluded data, resulting in superior performance on occluded person re-identification tasks.

What carries the argument

Teacher-student framework with co-saliency network (backbone features feeding into classification and co-saliency branches) and cross-domain simulator (artificial occlusions added with increasing probability).

If this is right

  • The framework reduces reliance on large quantities of occluded training data.
  • Automatic saliency guidance allows the model to focus on visible body parts without labels.
  • Progressive occlusion simulation helps bridge the domain gap between full-body and occluded images.
  • Outperforms prior methods on four standard occluded re-id benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could extend to other vision tasks with domain shifts due to missing data, such as occluded object recognition.
  • Removing the need for manual annotations in saliency might inspire similar techniques in unsupervised learning settings.
  • If the simulator's growing probability is key, varying the schedule could be tested for optimal transfer.

Load-bearing premise

The co-saliency network guides the model to highlight meaningful parts without manual annotation and the cross-domain simulator bridges the full-body to occluded domain gap.

What would settle it

If the proposed method does not achieve higher accuracy than state-of-the-art methods on the four occluded person re-id benchmarks, the effectiveness of the framework would be disproven.

Figures

Figures reproduced from arXiv: 1907.03253 by Jianhuang Lai, Jiaxuan Zhuo, Peijia Chen.

Figure 1
Figure 1. Figure 1: Scenario of occluded person re-id and samples [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Motivation of our proposal. The histogram repre [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our proposed framework. The teacher network learns a basic occlusion-robust model only using large [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of saliency maps for real-world oc [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance comparisons of the cross-domain [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Samples of existing salient object detectors and our [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: CMC curves of key components. Dashed lines represent the performances without the "student" stage (The left [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparisons with state-of-the-art on CMC curve. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
read the original abstract

Person re-identification (re-id) has made great progress in recent years, but occlusion is still a challenging problem which significantly degenerates the identification performance. In this paper, we design a teacher-student learning framework to learn an occlusion-robust model from the full-body person domain to the occluded person domain. Notably, the teacher network only uses large-scale full-body person data to simulate the learning process of occluded person re-id. Based on the teacher network, the student network then trains a better model by using inadequate real-world occluded person data. In order to transfer more knowledge from the teacher network to the student network, we equip the proposed framework with a co-saliency network and a cross-domain simulator. The co-saliency network extracts the backbone features, and two separated collaborative branches are followed by the backbone. One branch is a classification branch for identity recognition and the other is a co-saliency branch for guiding the network to highlight meaningful parts without any manual annotation. The cross-domain simulator generates artificial occlusions on full-body person data under a growing probability so that the teacher network could train a cross-domain model by observing more and more occluded cases. Experiments on four occluded person re-id benchmarks show that our method outperforms other state-of-the-art methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a teacher-student framework for occluded person re-identification. The teacher network is trained solely on large-scale full-body data augmented by a cross-domain simulator that introduces artificial occlusions with growing probability; it incorporates a co-saliency network (backbone plus collaborative classification and co-saliency branches) to highlight identity-relevant regions without manual supervision. The student network then refines the model using limited real occluded data. The central empirical claim is that the resulting model outperforms prior state-of-the-art methods on four occluded re-id benchmarks.

Significance. If the reported gains are reproducible, the work would be significant because it demonstrates a practical, annotation-free transfer mechanism that leverages abundant full-body datasets to improve robustness on occluded domains. The combination of growing-occlusion simulation and unsupervised co-saliency provides a concrete recipe that could be adopted or extended by the re-id community.

major comments (2)
  1. [§3.2] §3.2 (cross-domain simulator): the description states that occlusion probability 'grows' during training, but no schedule, functional form, or ablation on the growth rate is supplied; without this, it is impossible to determine whether the reported gains depend on a specific curriculum or would hold under a constant probability.
  2. [§4] §4 (experiments): the abstract asserts outperformance on four benchmarks, yet the provided text supplies neither quantitative tables, baseline numbers, nor statistical significance tests; the central claim therefore cannot be evaluated for effect size or robustness.
minor comments (2)
  1. [§3.1] The co-saliency branch is described as 'collaborative' yet the loss combination and gradient flow between the classification and co-saliency heads are not formalized; an equation or diagram would clarify the training objective.
  2. [§4] Dataset statistics (number of identities, images, occlusion ratios) for the four benchmarks are referenced but not tabulated; adding a summary table would aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (cross-domain simulator): the description states that occlusion probability 'grows' during training, but no schedule, functional form, or ablation on the growth rate is supplied; without this, it is impossible to determine whether the reported gains depend on a specific curriculum or would hold under a constant probability.

    Authors: We acknowledge that the current description lacks sufficient detail on the occlusion probability schedule. In the revised version, we will provide the exact functional form (e.g., a linear growth schedule from 0 to 0.5) and conduct an ablation study on the growth rate to show its impact. revision: yes

  2. Referee: [§4] §4 (experiments): the abstract asserts outperformance on four benchmarks, yet the provided text supplies neither quantitative tables, baseline numbers, nor statistical significance tests; the central claim therefore cannot be evaluated for effect size or robustness.

    Authors: The manuscript's experimental section does contain quantitative comparisons on the four benchmarks. To ensure clarity, we will highlight the tables more prominently and add statistical significance tests in the revision. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper presents a teacher-student framework with a co-saliency network and cross-domain simulator to transfer knowledge from full-body to occluded re-id domains. The central claim is empirical outperformance on four external benchmarks, which is directly falsifiable against independent test sets and does not reduce to any fitted parameter, self-citation chain, or definitional equivalence. No equations, ansatzes, or uniqueness theorems are invoked in the provided text; the method description is a standard architectural proposal whose validity rests on external validation rather than internal construction. This matches the default expectation of no circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no equations, training details, or implementation specifics, so no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.0 · 5759 in / 993 out tokens · 21125 ms · 2026-05-25T01:32:43.224391+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 5 internal anchors

  1. [1]

    Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Süsstrunk

    work page

  2. [2]

    Frequency-tuned salient region detection. In CVPR. 1597–1604

    work page

  3. [3]

    Xiaobin Chang, Timothy M Hospedales, and Tao Xiang. 2018. Multi-level fac- torisation net for person re-identification. In CVPR. 2109–2118. Table 5: Comparisons with state-of-the-art on rank-1/5. Dataset Occluded-REID Partial-REID P-DukeMTMC P-ETHZ Methods r=1 r=5 r=1 r=5 r=1 r=5 r=1 r=5 XQDA [15] 36.71 65.11 33.14 66.18 15.93 27.50 44.98 70.88 GOG [19] 4...

    work page 2018

  4. [4]

    Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. 2017. Beyond triplet loss: a deep quadruplet network for person re-identification. (2017), 403– 412

    work page 2017

  5. [5]

    De Cheng, Yihong Gong, Sanping Zhou, Jinjun Wang, and Nanning Zheng. 2016. Person re-identification by multi-channel parts-based cnn with improved triplet loss function. In CVPR. 1335–1344

    work page 2016

  6. [6]

    Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. 2008. A mobile vision system for robust multi-person tracking. (2008), 1–8

    work page 2008

  7. [7]

    Xing Fan, Hao Luo, Xuan Zhang, Lingxiao He, Chi Zhang, and Wei Jiang. 2018. SCPNet: Spatial-Channel Parallelism Network for Joint Holistic and Partial Person Re-Identification. arXiv preprint arXiv:1810.06996 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  8. [8]

    Douglas Gray, Shane Brennan, and Hai Tao. 2007. Evaluating appearance models for recognition, reacquisition, and tracking. In PETS, Vol. 3. Citeseer, 1–7. 8

    work page 2007

  9. [9]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In CVPR. 770–778

    work page 2016

  10. [10]

    Lingxiao He, Jian Liang, Haiqing Li, and Zhenan Sun. 2018. Deep spatial feature reconstruction for partial person re-identification: Alignment-free approach. In CVPR. 7073–7082

    work page 2018

  11. [11]

    Alexander Hermans, Lucas Beyer, and Bastian Leibe. 2017. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  12. [12]

    Qibin Hou, Ming-Ming Cheng, Xiaowei Hu, Ali Borji, Zhuowen Tu, and Philip HS Torr. 2017. Deeply supervised salient object detection with short connections. In CVPR. 3203–3212

    work page 2017

  13. [13]

    Mahdi M Kalayeh, Emrah Basaran, Muhittin Gökmen, Mustafa E Kamasak, and Mubarak Shah. 2018. Human semantic parsing for person re-identification. In CVPR. 1062–1071

    work page 2018

  14. [14]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  15. [15]

    Dangwei Li, Xiaotang Chen, Zhang Zhang, and Kaiqi Huang. 2017. Learning deep context-aware features over body and latent parts for person re-identification. (2017), 384–393

    work page 2017

  16. [16]

    Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z Li. 2015. Person re- identification by local maximal occurrence representation and metric learning. In CVPR. 2197–2206

    work page 2015

  17. [17]

    Nian Liu, Junwei Han, and Ming-Hsuan Yang. 2018. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR. 3089–3098

    work page 2018

  18. [18]

    Xihui Liu, Haiyu Zhao, Maoqing Tian, Lu Sheng, Jing Shao, Shuai Yi, Junjie Yan, and Xiaogang Wang. 2017. Hydraplus-net: Attentive deep features for pedestrian analysis. In ICCV. 350–359

    work page 2017

  19. [19]

    Zhiming Luo, Akshaya Mishra, Andrew Achkar, Justin Eichel, Shaozi Li, and Pierre-Marc Jodoin. 2017. Non-local deep features for salient object detection. In CVPR. 6609–6617

    work page 2017

  20. [20]

    Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. 2016. Hier- archical gaussian descriptor for person re-identification. In CVPR. 1363–1372

    work page 2016

  21. [21]

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer

    work page

  22. [22]

    Automatic differentiation in pytorch. (2017)

    work page 2017

  23. [23]

    Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi

    work page

  24. [24]

    Performance measures and a data set for multi-target, multi-camera track- ing. In ECCV. Springer, 17–35

    work page

  25. [25]

    Yifan Sun, Liang Zheng, Weijian Deng, and Shengjin Wang. 2017. Svdnet for pedestrian retrieval. (2017), 3800–3808

    work page 2017

  26. [26]

    Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. 2018. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV. 480–496

    work page 2018

  27. [27]

    Rahul Rama Varior, Mrinal Haloi, and Gang Wang. 2016. Gated siamese convolu- tional neural network architecture for human re-identification. (2016), 791–808

    work page 2016

  28. [28]

    Tong Xiao, Hongsheng Li, Wanli Ouyang, and Xiaogang Wang. 2016. Learn- ing deep feature representations with domain guided dropout for person re- identification. In CVPR. 1249–1258

    work page 2016

  29. [29]

    Qian Yu, Xiaobin Chang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales

    work page

  30. [30]

    The Devil is in the Middle: Exploiting Mid-level Representations for Cross-Domain Instance Matching

    The Devil is in the Middle: Exploiting Mid-level Representations for Cross- Domain Instance Matching. arXiv preprint arXiv:1711.08106 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  31. [31]

    Li Zhang, Tao Xiang, and Shaogang Gong. 2016. Learning a discriminative null space for person re-identification. In CVPR. 1239–1248

    work page 2016

  32. [32]

    Shifeng Zhang, Longyin Wen, Xiao Bian, Zhen Lei, and Stan Z Li. 2018. Occlusion- aware R-CNN: detecting pedestrians in a crowd. In ECCV. 637–653

    work page 2018

  33. [33]

    Liang Zheng, Zhi Bie, Yifan Sun, Jingdong Wang, Chi Su, Shengjin Wang, and Qi Tian. 2016. MARS: A Video Benchmark for Large-Scale Person Re-identification. In ECCV. Springer

    work page 2016

  34. [36]

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian

    work page

  35. [37]

    Scalable person re-identification: A benchmark. In ICCV. 1116–1124

    work page

  36. [38]

    Wei Shi Zheng, Li Xiang, Xiang Tao, Shengcai Liao, Jianhuang Lai, and Shaogang Gong. 2016. Partial Person Re-Identification. In ICCV

    work page 2016

  37. [39]

    Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. 2017. Random erasing data augmentation. arXiv preprint arXiv:1708.04896 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  38. [40]

    Chunluan Zhou and Junsong Yuan. 2018. Bi-box Regression for Pedestrian Detection and Occlusion Estimation. In ECCV. 135–151

    work page 2018

  39. [41]

    Zheng Zhu, Qiang Wang, Bo Li, Wei Wu, Junjie Yan, and Weiming Hu. 2018. Distractor-aware siamese networks for visual object tracking. InECCV. 101–117

    work page 2018

  40. [42]

    Jiaxuan Zhuo, Zeyu Chen, Jianhuang Lai, and Guangcong Wang. 2018. Occluded Person Re-identification. (2018). 9

    work page 2018