Back

Beyond Part Models: Person Retrieval with Refined Part Pooling

Yifan SunLiang ZhengYi YangQi TianShengjin Wang

article2017ECCV2,428 citations

Proposes a part-based convolutional baseline and a refined part pooling method that dynamically reassigns misaligned feature outliers across uniform horizontal partitions, setting a high standard for person re-identification without requiring external pose estimation.

Listen

Person retrieval, or person re-identification, is critical for public safety, security surveillance, and smart city systems that track individuals across multi-camera networks. While deep learning methods using part-level visual features provide strong accuracy, existing techniques generally rely on complex external pose estimators—which suffer from dataset mismatches—or unguided attention mechanisms that struggle to localize body parts consistently.

The article demonstrates that high-accuracy person retrieval can be achieved without external pose estimators or human part labels by combining a uniform convolutional partitioning baseline with an adaptive refinement method that reinforces visual consistency within each body part.

To evaluate this framework, the authors developed the Part-based Convolutional Baseline (PCB), which divides high-level convolutional feature maps into uniform horizontal stripes and applies separate classification losses to each part. To resolve misalignments caused by rigid horizontal slicing, they introduced Refined Part Pooling (RPP), a weakly supervised module that adaptively reassigns outlier visual features to their closest corresponding body parts. The framework was evaluated across three standard person retrieval benchmarks: Market-1501 (1,501 identities across 6 cameras), DukeMTMC-reID (1,404 identities across 8 cameras), and CUHK03 (1,467 identities across 2 cameras).

The experimental findings demonstrate significant performance gains. First, the baseline PCB architecture alone established state-of-the-art results across all benchmarks, lifting mean Average Precision (mAP)—a measure of comprehensive retrieval quality—from 68.5% to 77.4% on Market-1501 and from 38.9% to 54.2% on CUHK03 compared to standard global baselines. Second, adding RPP further boosted accuracy across all datasets, delivering 93.8% Rank-1 accuracy and 81.6% mAP on Market-1501, 83.3% Rank-1 and 69.2% mAP on DukeMTMC-reID, and 63.7% Rank-1 and 57.5% mAP on CUHK03 without post-processing re-ranking. Third, architectural ablations proved that learning distinct part classifiers with independent parameters is vital, outperforming shared-parameter and single-loss alternatives by 2.4 to 7.4 percentage points. Finally, hyperparameter evaluations identified that partitioning feature maps into six parts provides the optimal granularity, whereas excessive slicing (e.g., eight or twelve parts) degrades accuracy due to redundant or collapsed part definitions.

These results show that engineering teams do not need expensive human part annotations or computationally intensive pose estimation networks to build high-performing retrieval systems. Instead, uniform partitioning coupled with weakly supervised part refinement delivers superior matching accuracy for challenging queries with minimal added computational cost (requiring only ten additional training epochs). In production surveillance and tracking deployments, this approach reduces computational overhead and annotation costs while improving identification reliability across camera handoffs.

Organizations implementing automated person retrieval should adopt the PCB architecture with six horizontal partitions and integrate RPP refinement. For high-precision production environments, operators can optionally combine this framework with post-processing re-ranking algorithms, which further elevated Market-1501 performance to 95.1% Rank-1 accuracy and 91.9% mAP in testing.

Confidence in these findings is high across standard single-query benchmarks. However, stakeholders should note that the evaluation was conducted on static image crops from standard research datasets. Real-world deployment will require further validation under severe occlusions, extreme lighting changes, and crowded camera angles where bounding-box quality varies significantly.

Australian National UniversityHuaweiTsinghua UniversityUniversity of Technology SydneyUniversity of Texas at San Antoniodescriptorspose estimationPartPCBPerson
arXiv: 1711.09349syafterzy/PCB_RPP
  • Paper: Object Detection with Discriminatively Trained Part-Based Models, Pedro F. Felzenszwalb et al. (2010). Understanding part-based models and latent SVMs is crucial because the source paper builds directly upon part-based feature representations for person retrieval.
  • Paper: Fast R-CNN, Ross B. Girshick (2015). Familiarity with Fast R-CNN provides the foundational convolutional feature-pooling concepts that the source paper adapts and refines for part-based pedestrian retrieval.
Cover for Beyond Part Models: Person Retrieval with Refined Part Pooling

Abstract

Employing part-level features for pedestrian image description offers fine-grained information and has been verified as beneficial for person retrieval in very recent literature. A prerequisite of part discovery is that each part should be well located. Instead of using external cues, e.g., pose estimation, to directly locate parts, this paper lays emphasis on the content consistency within each part.

Specifically, we target at learning discriminative part-informed features for person retrieval and make two contributions. (i) A network named Part-based Convolutional Baseline (PCB). Given an image input, it outputs a convolutional descriptor consisting of several part-level features. With a uniform partition strategy, PCB achieves competitive results with the state-of-the-art methods, proving itself as a strong convolutional baseline for person retrieval.

(ii) A refined part pooling (RPP) method. Uniform partition inevitably incurs outliers in each part, which are in fact more similar to other parts. RPP re-assigns these outliers to the parts they are closest to, resulting in refined parts with enhanced within-part consistency. Experiment confirms that RPP allows PCB to gain another round of performance boost. For instance, on the Market-1501 dataset, we achieve (77.4+4.2)% mAP and (92.3+1.5)% rank-1 accuracy, surpassing the state of the art by a large margin.

Topics

descriptorspose estimation

Keywords

PartPCBPersonPart PoolingPerson RetrievalNed Part PoolingBeyond Part ModelsRPPPart ClassiPartition

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 3 PCB: A Strong Convolutional Baseline
  • 3.1 Structure of PCB
  • 3.2 Important Parameters.
  • 3.3 Potential Alternative Structures
  • 4 Refined Part Pooling
  • 4.1 Within-Part Inconsistency
  • 4.2 Relocating Outliers
  • 4.3 Induced Training for Part Classifier
  • 4.4 Discussions on Refined Part Pooling
  • 5 Experiments
  • 5.1 Datasets and Settings
  • 5.2 Implementation details
  • 5.3 Performance evaluation
  • 5.4 Parameters Analysis
  • 5.5 Induction and Attention Mechanism
  • 6 Conclusion
  • References

Knowls

  1. Knowl 1 — Part-Based Convolutional Baseline Architecture for Person Re-Identification

    model/method

    The Part-based Convolutional Baseline (PCB) is a deep neural network architecture for person re-identification that extracts fine-grained, part-level spatial descriptors from convolutional activation tensors without requiring explicit image slicing or auxiliary pose estimation.

    Given an input pedestrian image of size H×W×3H \times W \times 3 (typically 384×128×3384 \times 128 \times 3), the network processes it through a convolutional backbone (e.g., ResNet-50 modified by removing the final global average pooling layer and removing the spatial down-sampling stride in the last stage Conv5, producing an activation tensor T∈RM×N×C\mathbf{T} \in \mathbb{R}^{M \times N \times C} with spatial size M×N=24×8M \times N = 24 \times 8 and C=2048C = 2048).

    PCB divides the tensor T\mathbf{T} uniformly along the vertical axis into pp non-overlapping horizontal stripes (default p=6p = 6). A conventional average pooling operation computes a part-level column vector for each stripe i∈{1,…,p}i \in \{1, \dots, p\}: gi=1∣Si∣∑f∈Sif\mathbf{g}_i = \frac{1}{|S_i|} \sum_{\mathbf{f} \in S_i} \mathbf{f} where SiS_i is the set of column vectors f∈RC\mathbf{f} \in \mathbb{R}^C within the ii-th horizontal stripe, yielding gi∈R2048\mathbf{g}_i \in \mathbb{R}^{2048}.

    Each column vector gi\mathbf{g}_i is passed through a 1×11 \times 1 convolutional layer to reduce its dimension to hi∈R256\mathbf{h}_i \in \mathbb{R}^{256}. Each hi\mathbf{h}_i is then input into an independent fully-connected (FC) layer followed by a softmax function to predict the identity of the person. The overall training loss is the sum of pp individual cross-entropy losses: Ltotal=∑i=1pLCE(y^i,y)\mathcal{L}_{\text{total}} = \sum_{i=1}^p \mathcal{L}_{\text{CE}}(\hat{y}_i, y) where yy is the ground-truth identity label and y^i\hat{y}_i is the predicted identity distribution from the ii-th part classifier. Crucially, the FC classification branches do not share parameters.

    For feature extraction during inference, the final pedestrian descriptor is formed by concatenating either the unreduced vectors G=[g1,g2,…,gp]\mathcal{G} = [\mathbf{g}_1, \mathbf{g}_2, \dots, \mathbf{g}_p] (12,288 dimensions for p=6p=6) or the dimension-reduced vectors H=[h1,h2,…,hp]\mathcal{H} = [\mathbf{h}_1, \mathbf{h}_2, \dots, \mathbf{h}_p] (1,536 dimensions for p=6p=6).

  2. Knowl 2 — Refined Part Pooling Mechanism for Adaptive Part Discovery

    model/method

    Refined Part Pooling (RPP) is an adaptive soft-pooling mechanism that addresses within-part inconsistency caused by rigid uniform spatial slicing. Uniform partitioning assumes spatial alignment across images, but variations in pose and viewpoint cause some spatial locations within a stripe to visually belong to other body regions (acting as outliers).

    RPP replaces hard uniform slicing with a soft, content-based assignment of spatial column vectors to parts. Let F={f(y,x)∈RC∣1≤y≤M,1≤x≤N}F = \{\mathbf{f}_{(y,x)} \in \mathbb{R}^C \mid 1 \le y \le M, 1 \le x \le N\} be the set of all spatial column vectors in the 3D feature tensor T∈RM×N×C\mathbf{T} \in \mathbb{R}^{M \times N \times C}. A part classifier with parameter weights W=[W1,W2,…,Wp]T∈Rp×C\mathbf{W} = [\mathbf{W}_1, \mathbf{W}_2, \dots, \mathbf{W}_p]^T \in \mathbb{R}^{p \times C} predicts the soft assignment probability S(f↔Pi)S(\mathbf{f} \leftrightarrow P_i) of each column vector f\mathbf{f} belonging to part PiP_i (i∈{1,…,p}i \in \{1, \dots, p\}): S(f↔Pi)=exp⁡(WiTf)∑j=1pexp⁡(WjTf)S(\mathbf{f} \leftrightarrow P_i) = \frac{\exp(\mathbf{W}_i^T \mathbf{f})}{\sum_{j=1}^p \exp(\mathbf{W}_j^T \mathbf{f})}

    Each refined part-level feature vector gi∈RC\mathbf{g}_i \in \mathbb{R}^C is then computed via weighted global average pooling across all spatial column vectors in FF: gi=∑f∈FS(f↔Pi)f∑f∈FS(f↔Pi)\mathbf{g}_i = \frac{\sum_{\mathbf{f} \in F} S(\mathbf{f} \leftrightarrow P_i) \mathbf{f}}{\sum_{\mathbf{f} \in F} S(\mathbf{f} \leftrightarrow P_i)}

    RPP enables dynamic, data-driven spatial aggregation where column vectors with similar visual semantics are pooled into the corresponding part feature regardless of slight spatial misalignments.

  3. Knowl 3 — Induced Training Procedure for the Part Classifier in Refined Part Pooling

    algorithm

    To train the part classifier weights W\mathbf{W} of Refined Part Pooling without requiring ground-truth human body part annotations, an induced training procedure leverages the representations learned by a pre-trained Part-based Convolutional Baseline (PCB).

    Input: Training dataset D={(Ik,yk)}D = \{(I_k, y_k)\}, backbone network, number of parts pp (e.g., p=6p=6), base learning rate η0=0.1\eta_0 = 0.1, fine-tuning learning rate η1=0.01\eta_1 = 0.01
    Output: Trained PCB+RPP model parameters
    Step 1: Train standard PCB to convergence
    Initialize backbone on ImageNet
    Train standard PCB with uniform spatial partitioning for 60 epochs using batch size 64 and initial learning rate η0\eta_0 (decayed to 0.1η00.1 \eta_0 after 40 epochs)
    Step 2: Append part classifier
    Remove uniform spatial average pooling
    Append a linear part classifier layer W∈Rp×C\mathbf{W} \in \mathbb{R}^{p \times C} on feature tensor T\mathbf{T}
    Configure soft pooling where each part PiP_i aggregates column vectors weighted by S(f↔Pi)=softmax(WiTf)S(\mathbf{f} \leftrightarrow P_i) = \text{softmax}(\mathbf{W}_i^T \mathbf{f})
    Step 3: Induced training of part classifier
    Freeze all pre-trained layers of PCB (backbone, dimension-reduction conv layers, and identity classification FC layers)
    Set only the part classifier W\mathbf{W} as trainable
    Train the model on dataset DD with learning rate η1\eta_1 until convergence (approximately 5 epochs)
    Step 4: Full network fine-tuning
    Unfreeze all layers in the network
    Fine-tune the entire model (backbone, part classifier, dimension-reduction conv, and identity FC classifiers) jointly on dataset DD with learning rate η1\eta_1 for 5 epochs
    return Optimized PCB+RPP network parameters

    In Step 3, because the frozen identity classifiers expect feature parts roughly corresponding to horizontal stripes, the part classifier is constrained toward the uniform partition while simultaneously grouping content-consistent column vectors, preventing degenerate solutions without external part annotations.

  4. Knowl 4 — Within-Part Inconsistency in Spatial Part Partitioning

    definition

    Within-part inconsistency refers to the phenomenon where spatial feature vectors grouped into the same partition by uniform horizontal slicing exhibit lower semantic similarity to their own part summary than to other parts.

    Formally, let T∈RM×N×C\mathbf{T} \in \mathbb{R}^{M \times N \times C} be a convolutional feature tensor uniformly partitioned along the height dimension into pp horizontal stripes, and let gi∈RC\mathbf{g}_i \in \mathbb{R}^C denote the average-pooled feature vector of stripe i∈{1,…,p}i \in \{1, \dots, p\}. For an individual column vector f∈RC\mathbf{f} \in \mathbb{R}^C physically located in stripe ii, within-part inconsistency occurs when: ∃j≠isuch thatf⋅gj∥f∥2∥gj∥2>f⋅gi∥f∥2∥gi∥2\exists j \neq i \quad \text{such that} \quad \frac{\mathbf{f} \cdot \mathbf{g}_j}{\|\mathbf{f}\|_2 \|\mathbf{g}_j\|_2} > \frac{\mathbf{f} \cdot \mathbf{g}_i}{\|\mathbf{f}\|_2 \|\mathbf{g}_i\|_2}

    Such column vectors represent spatial partition outliers caused by human pose variation, non-rigid body deformation, or background clutter, degrading the discriminative capability of part-level representations unless adaptively reassigned.

  5. Knowl 5 — Performance Comparison of PCB Architecture Variants and Baseline Models

    data/table

    An empirical evaluation on three person re-identification benchmarks (Market-1501, DukeMTMC-reID, and CUHK03 under the 767/700 split protocol) demonstrates the impact of PCB's design choices:

    Models Feature dim Market-1501 DukeMTMC-reID CUHK03
    R-1 (%) mAP (%) R-1 (%) mAP (%) R-1 (%) mAP (%)
    IDE (pool5) 2048 85.3 68.5 73.2 52.8 43.8 38.9
    IDE (FC) 256 83.8 67.7 72.4 51.6 43.3 38.3
    Variant 1 (G\mathcal{G}) 12288 86.7 69.4 73.9 53.2 43.6 38.8
    Variant 1 (H\mathcal{H}) 1536 85.6 68.3 72.8 52.5 44.1 39.1
    Variant 2 (G\mathcal{G}) 12288 91.2 75.0 80.2 62.8 52.6 45.8
    Variant 2 (H\mathcal{H}) 1536 91.0 75.3 80.0 62.6 54.0 47.2
    PCB (G\mathcal{G}) 12288 92.3 77.4 81.7 66.1 59.7 53.2
    PCB (H\mathcal{H}) 1536 92.4 77.3 81.9 65.3 61.3 54.2
    PCB+RPP (G\mathcal{G}) 12288 93.8 81.6 83.3 69.2 62.8 56.7
    PCB+RPP (H\mathcal{H}) 1536 93.1 81.0 82.9 68.5 63.7 57.5

    Key observations:

    1. PCB vs IDE baseline: PCB significantly outperforms IDE across all datasets (+7.0% Rank-1 and +8.9% mAP on Market-1501; +8.5% Rank-1 and +13.3% mAP on DukeMTMC-reID; +17.5% Rank-1 and +15.3% mAP on CUHK03), showing the advantage of part-informed deep descriptors over global representations.
    2. Independent vs Single Loss (Variant 1): Averaging part vectors into a single vector trained with one loss yields much lower accuracy than PCB (69.4% vs 77.4% mAP on Market-1501), confirming that independent per-part losses are essential.
    3. Parameter Sharing (Variant 2): Sharing FC classifier weights across all stripes underperforms unshared classifiers by 2.4% to 7.4% in Rank-1 accuracy.
    4. RPP Boost: RPP brings further improvement over PCB on all datasets, particularly in mAP (+4.2% on Market-1501, +3.1% on DukeMTMC-reID, +3.5% on CUHK03 for G\mathcal{G}).
    5. Unreduced (G\mathcal{G}) vs Reduced (H\mathcal{H}) Descriptors: Concatenating unreduced 2048-d vectors G\mathcal{G} yields slightly higher mAP, while dimension-reduced 256-d vectors H\mathcal{H} achieve competitive performance with an 8×8\times lower dimensionality.

  6. Knowl 6 — Person Re-Identification Performance Comparison Across Benchmark Datasets

    data/table

    PCB and PCB+RPP achieve state-of-the-art retrieval performance across Market-1501, DukeMTMC-reID, and CUHK03 (DPM-detected bounding boxes, new split protocol) under single-query mode without re-ranking:

    Method Market-1501 DukeMTMC-reID CUHK03
    Rank-1 (%) mAP (%) Rank-1 (%) mAP (%) Rank-1 (%) mAP (%)
    BoW+kissme 44.4 20.8 25.1 12.2 6.4 6.4
    SVDNet 82.3 62.1 76.7 56.8 41.5 37.3
    PAR 81.0 63.4 - - - -
    PDC (with pose labels) 84.4 63.4 - - - -
    GLAD (with pose labels) 89.9 73.9 - - - -
    HA-CNN 91.2 75.7 80.5 63.8 41.7 38.6
    MultiScale 88.9 73.1 79.2 60.6 40.7 37.0
    PCB (Uniform Partition) 92.3 77.4 81.8 66.1 61.3 54.2
    PCB+RPP (Refined Part Pooling) 93.8 81.6 83.3 69.2 63.7 57.5

    Without using external pose estimation models or human part annotations, PCB outperforms existing methods that rely on auxiliary pose labels (e.g., PDC and GLAD). Adding RPP provides an additional boost across all metrics. When combined with kk-reciprocal re-ranking on Market-1501, PCB+RPP reaches 95.1% Rank-1 accuracy and 91.9% mAP.

  7. Knowl 7 — Impact of Spatial Down-Sampling Rate and Image Resolution on Feature Granularity

    empirical result

    In standard classification backbones like ResNet-50, the spatial down-sampling stride in the Conv5 block produces coarse spatial dimensions (12×412 \times 4 for an input of 384×128384 \times 128). Removing the spatial down-sampling stride in Conv5 increases the output tensor T\mathbf{T} spatial dimensions by 4×4\times to 24×824 \times 8.

    Empirical evaluation on Market-1501 across image resolutions ranging from 192×64192 \times 64 to 576×192576 \times 192 (in steps of 96×3296 \times 32) shows:

    1. Resolution Impact: Performance improves with image size up to 384×128384 \times 128, beyond which Rank-1 accuracy and mAP plateau.
    2. Granularity Enhancement: Halving the backbone down-sampling rate consistently boosts retrieval accuracy. PCB operating on 384×128384 \times 128 images with the halved down-sampling rate matches the accuracy of PCB on 576×192576 \times 192 images with the standard down-sampling rate, while requiring lower overall computation.

  8. Knowl 8 — Impact of Horizontal Part Count on Re-Identification Accuracy and Part Collapse

    empirical result

    The number of horizontal parts pp determines the granularity of part representations in PCB and PCB+RPP. Testing values p∈{1,2,4,6,8,12}p \in \{1, 2, 4, 6, 8, 12\} on Market-1501 demonstrates:

    1. Accuracy Peak: Setting p=1p=1 yields a global descriptor with the lowest accuracy. Both mAP and Rank-1 accuracy increase as pp grows from 1 to 6, peaking at p=6p=6 for both uniform partition PCB (77.4% mAP, 92.3% Rank-1) and PCB+RPP (81.6% mAP, 93.8% Rank-1).
    2. Degeneracy at Higher pp: Increasing pp to 8 or 12 causes a sharp decline in performance. Feature visualizations show that at p≥8p \ge 8, several parts learn redundant weights identical to neighboring parts, or collapse entirely into empty parts with negligible activation. Consequently, p=6p=6 provides the optimal partition granularity.

  9. Knowl 9 — Ablation of Induced Training Against Unsupervised Attention and Mid-Level Element Mining

    empirical result

    The induced training procedure for PCB+RPP was ablated on Market-1501 and DukeMTMC-reID against attention-driven soft partition and pseudo-label mid-level element mining:

    Methods Market-1501 DukeMTMC-reID
    Rank-1 (%) mAP (%) Rank-1 (%) mAP (%)
    PAR 81.0 63.4 - -
    IDE baseline 85.3 68.5 73.2 52.8
    RPP (w/o induction) 88.7 74.6 78.8 60.9
    PCB (uniform partition) 92.3 77.4 81.7 66.1
    PCB+RPP 93.8 81.6 83.3 69.2

    Key empirical findings:

    1. Necessity of Induction: Training the part classifier from scratch without pre-trained PCB induction ("RPP w/o induction") yields 74.6% mAP on Market-1501 and 60.9% on DukeMTMC-reID. While superior to the global IDE baseline (68.5% and 52.8% mAP), it significantly underperforms uniform PCB (77.4% and 66.1% mAP) and induced PCB+RPP (81.6% and 69.2% mAP).
    2. Comparison to Mid-Level Mining: Training the part classifier using discrete pseudo part labels assigned to each horizontal stripe achieves 93.0% Rank-1 / 79.0% mAP on Market-1501 and 82.1% Rank-1 / 66.9% mAP on DukeMTMC-reID. While this improves on uniform PCB, it remains inferior to RPP (81.6% and 69.2% mAP), because RPP's part classifier is optimized jointly with identity classification rather than artificial pseudo labels.

Coverage note — None of the substantial contributed material was omitted.

References

  1. 1.Barbosa, I.B., Cristani, M., Caputo, B., Rognhaugen, A., Theoharis, T.: Looking beyond appearances: Synthetic training data for deep cnns in re-identification. arXiv preprint arXiv:1701.03153 (2017)
  2. 2.Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2D pose estimation using part affinity fields. In: CVPR (2017)
  3. 3.Chen, Y., Zhu, X., Gong, S.: Person re-identification by deep learning multi-scale representations. In: International Conference on Computer Vision, Workshop on Cross-Domain Human Identification (CHI) (2017)
  4. 4.Cheng, D.S., Cristani, M., Stoppa, M., Bazzani, L., Murino, V.: Custom pictorial structures for re-identification. In: BMVC (2011)
  5. 5.Dai, J., Li, Y., He, K., Sun, J.: R-FCN: object detection via region-based fully convolutional networks. In: NIPS (2016)
  6. 6.Das, A., Chakraborty, A., Roy-Chowdhury, A.K.: Consistent re-identification in a camera network. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014. LNCS, vol. 8690, pp. 330–345. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-10605-2_22
  7. 7.Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Li, F.F.: Imagenet: a large-scale hierarchical image database. In: CVPR (2009)
  8. 8.Diba, A., Pazandeh, A.M., Pirsiavash, H., Gool, L.V.: Deepcamp: deep convolutional action & attribute mid-level patterns. In: CVPR (2016)
  9. 9.Engel, C., Baumgartner, P., Holzmann, M., Nutzel, J.F.: Person re-identification by support vector ranking. In: BMVC (2010)
  10. 10.Felzenszwalb, P., McAllester, D., Ramanan, D.: A discriminatively trained, multiscale, deformable part model. In: CVPR (2008)
  11. 11.Geng, M., Wang, Y., Xiang, T., Tian, Y.: Deep transfer learning for person re-identification. arXiv preprint arXiv:1611.05244 (2016)
  12. 12.Gheissari, N., Sebastian, T.B., Hartley, R.: Person reidentification using spatiotemporal appearance. In: CVPR (2006)
  13. 13.Gray, D., Tao, H.: Viewpoint invariant pedestrian recognition with an ensemble of localized features. In: Forsyth, D., Torr, P., Zisserman, A. (eds.) ECCV 2008. LNCS, vol. 5302, pp. 262–275. Springer, Heidelberg (2008). https://doi.org/10.1007/978-3-540-88682-2_21
  14. 14.He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)
  15. 15.Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv: 1703.07737 (2017)
  16. 16.Insafutdinov, E., Pishchulin, L., Andres, B., Andriluka, M., Schiele, B.: DeeperCut: a deeper, stronger, and faster multi-person pose estimation model. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9910, pp. 34–50. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46466-4_3
  17. 17.Jose, C., Fleuret, F.: Scalable metric learning via weighted approximate rank component analysis. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9909, pp. 875–890. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46454-1_53
  18. 18.Karanam, S., Gou, M., Wu, Z., Rates-Borras, A., Camps, O., Radke, R.J.: A comprehensive evaluation and benchmark for person re-identification: features, metrics, and datasets. arXiv preprint arXiv: 1605.09653 (2016)
  19. 19.Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: deep filter pairing neural network for person re-identification. In: CVPR (2014)
  20. 20.Li, W., Zhu, X., Gong, S.: Person re-identification by deep joint learning of multi-loss classification. In: IJCAI (2017)
  21. 21.Li, W., Zhu, X., Gong, S.: Harmonious attention network for person re-identification. arXiv preprint arXiv:1802.08122 (2018)
  22. 22.Li, Y., Liu, L., Shen, C., van den Hengel, A.: Mining mid-level visual patterns with deep CNN activations. Int. J. Comput. Vision (2017)
  23. 23.Liao, S., Hu, Y., Zhu, X., Li, S.Z.: Person re-identification by local maximal occurrence representation and metric learning. In: CVPR (2015)
  24. 24.Liu, W., et al.: SSD: single shot multibox detector. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9905, pp. 21–37. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46448-0_2
  25. 25.Liu, X., et al.: Hydraplus-net: attentive deep features for pedestrian analysis. In: ICCV (2017)
  26. 26.Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015)
  27. 27.M., J.O., Tuytelaars, T.: Modeling visual compatibility through hierarchical mid-level elements. In: ECCV (2016)
  28. 28.Ma, A.J., Yuen, P.C., Li, J.: Domain transfer support vector ranking for person re-identification without target camera label information. In: ICCV (2013)
  29. 29.Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9912, pp. 483–499. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46484-8_29
  30. 30.Ristani, E., Solera, F., Zou, R., Cucchiara, R., Tomasi, C.: Performance measures and a data set for multi-target, multi-camera tracking. In: Hua, G., J'egou, H. (eds.) ECCV 2016. LNCS, vol. 9914, pp. 17–35. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-48881-3_2
  31. 31.Su, C., Li, J., Zhang, S., Xing, J., Gao, W., Tian, Q.: Pose-driven deep convolutional model for person re-identification. In: ICCV (2017)
  32. 32.Sun, Y., Zheng, L., Deng, W., Wang, S.: SVDNet for pedestrian retrieval. In: ICCV (2017)
  33. 33.Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: AAAI (2017)
  34. 34.Ustinova, E., Ganin, Y., Lempitsky, V.: Multiregion bilinear convolutional neural networks for person re-identification. arXiv preprint arXiv: 1512.05300 (2015)
  35. 35.Wei, L., Zhang, S., Yao, H., Gao, W., Tian, Q.: GLAD: Global-local-alignment descriptor for pedestrian retrieval. ACM Multimed. (2017)
  36. 36.Wei, S.E., Ramakrishna, V., Kanade, T., Sheikh, Y.: Convolutional pose machines. In: CVPR (2016)
  37. 37.Xiao, T., Li, H., Ouyang, W., Wang, X.: Learning deep feature representations with domain guided dropout for person re-identification. In: CVPR (2016)
  38. 38.Xu, K., et al.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015)
  39. 39.Yao, H., Zhang, S., Zhang, Y., Li, J., Tian, Q.: Deep representation learning with part loss for person re-identification. arXiv preprint arXiv:1707.00798 (2017)
  40. 40.Zhang, Y., Xiang, T., Hospedales, T.M., Lu, H.: Deep mutual learning. arXiv preprint arXiv: 1705.00384 (2017)
  41. 41.Zhao, L., Li, X., Wang, J., Zhuang, Y.: Deeply-learned part-aligned representations for person re-identification. In: ICCV (2017)
  42. 42.Zheng, L., Huang, Y., Lu, H., Yang, Y.: Pose invariant embedding for deep person re-identification. arXiv preprint arXiv:1701.07732 (2017)
  43. 43.Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: a benchmark. In: ICCV (2015)
  44. 44.Zheng, L., Yang, Y., Hauptmann, A.G.: Person re-identification: past, present and future. arXiv preprint arXiv:1610.02984 (2016)
  45. 45.Zheng, W., Gong, S., Xiang, T.: Reidentification by relative distance comparison. TPAMI (2013)
  46. 46.Zheng, Z., Zheng, L., Yang, Y.: Pedestrian alignment network for large-scale person re-identification. arXiv preprint arXiv: 1707.00408 (2017)
  47. 47.Zheng, Z., Zheng, L., Yang, Y.: Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In: ICCV (2017)
  48. 48.Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking person re-identification with k-reciprocal encoding. In: CVPR (2017)
  49. 49.Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. arXiv preprint arXiv: 1708.04896 (2017)
  50. 50.Zhong, Z., Zheng, L., Zheng, Z., Li, S., Yang, Y.: Camera style adaptation for person re-identification. arXiv preprint arXiv:1711.10295 (2017)

Citation

MLA
Sun, Y., et al. “Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline)”. arXiv, 2017, http://arxiv.org/abs/1711.09349v3.
APA
Sun, Y., Zheng, L., Yang, Y., Tian, Q., & Wang, S. (2017). Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline). arXiv. http://arxiv.org/abs/1711.09349v3
Chicago
Sun, Y., L. Zheng, Y. Yang, Q. Tian, and S. Wang. 2017. “Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline)”. arXiv. http://arxiv.org/abs/1711.09349v3.
Harvard
Sun, Y. et al. (2017) “Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline)”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1711.09349v3.
Vancouver
1. Sun Y, Zheng L, Yang Y, Tian Q, Wang S (2017) Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline). arXiv

BibTeX

@article{sun2017beyond,
  title = {Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline)},
  author = {Sun, Yifan and Zheng, Liang and Yang, Yi and Tian, Qi and Wang, Shengjin},
  year = {2017},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1711.09349v3},
  eprint = {1711.09349}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF