Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function

De ChengYihong GongSanping ZhouJinjun WangN. Zheng

article2016CVPR1,327 citations

Presents a multi-channel convolutional architecture paired with an improved triplet loss that enforces an upper bound on positive pair distances, jointly learning global and body-part representations to improve person re-identification across non-overlapping camera views.

Listen

Identifying the same individual across multiple non-overlapping surveillance cameras remains a critical yet difficult task in automated security, robotics, and video analytics. The core challenges arise from severe variations in camera angles, lighting conditions, human body poses, and occlusions, combined with the fact that facial features are rarely clear enough for standard biometrics. Historically, systems tackled this by treating visual feature extraction and distance comparison as separate steps, which limited overall matching performance. The main objective of the article is to demonstrate an integrated deep learning approach that jointly learns both full-body and body-part visual features alongside an optimized distance metric using an enhanced loss function.

The authors evaluated their approach using standard experimental benchmarks on four public surveillance datasets: i-LIDS, VIPeR, PRID2011, and CUHK01. Their technical strategy uses a single multi-channel deep neural network that simultaneously processes a global full-body channel and four regional body-part channels with distinct filter sizes. This model was trained on image triplets—matching pairs versus non-matching individuals—using an improved triplet loss formulation. Unlike standard triplet loss, which only ensures that non-matching images are farther apart than matching ones, the improved loss explicitly enforces a compact margin on matching pairs, pulling features of the same individual much closer together.

The experimental findings show significant performance gains across all benchmark datasets. First, the proposed framework achieved top-ranked identification rates, outperforming leading traditional, deep learning, and metric ensemble methods across all four datasets. Second, incorporating body-part channels significantly boosted matching accuracy, providing up to a 13% improvement over full-body-only models. Third, replacing standard triplet loss with the improved loss function alone increased identification accuracy by up to 4%. Finally, detailed regional analysis revealed that the upper body (face and shoulders) provides the most discriminative and reliable features, while the lower body (legs and feet) contributes the least due to high movement variability.

These results imply that surveillance and retrieval systems can achieve substantially higher automated tracking accuracy without manual intervention or multi-stage pipelines. By capturing both holistic body appearance and fine-grained local parts within a single unified network, the model improves recognition reliability under extreme viewpoint and posture shifts. For operational decision-makers, this translates to reduced false-match rates and improved real-time tracking performance across disconnected camera networks.

Moving forward, stakeholders should consider adopting joint global-and-local feature learning frameworks when upgrading automated video surveillance infrastructure. The article suggests extending this multi-channel triplet architecture to broader multimedia tasks, such as large-scale image and video retrieval. Future work should validate performance on larger, unconstrained real-world deployments, as the current evaluation relies on standardized benchmark splits and requires tuning network depth and loss weighting parameters to match specific dataset scales.

Institute of Artificial Intelligence and Robotics Xi'an Jiaotong University convolutional neural networks feature space full body person re-identification multi-channel parts-based CNN triplet loss function

Cheng et al (2016).pdf

Paper: Deep Metric Learning Using Triplet Network, Elad Hoffer et al. (2014). It introduces the triplet network architecture and relative distance comparison framework that underpin the triplet loss formulations used for metric learning in person re-identification.
Paper: FaceNet: A unified embedding for face recognition and clustering, Florian Schroff et al. (2015). It establishes end-to-end deep metric learning with online hard-triplet loss optimization, which the source paper directly adapts and improves upon for multi-channel person re-identification.
Paper: Scalable Person Re-identification: A Benchmark, Liang Zheng et al. (2015). It establishes essential evaluation benchmarks and deep baseline practices for scalable person re-identification across non-overlapping camera views.
Paper: Person re-identification by Local Maximal Occurrence representation and metric learning, Shengcai Liao et al. (2014). It provides foundational insights into localized feature representations and cross-camera metric learning on standard benchmarks like VIPeR and CUHK01.
Paper: Learning a similarity metric discriminatively, with application to face verification, Sumit Chopra et al. (2005). It introduces the foundational concept of discriminative metric learning using Siamese convolutional neural networks for identity verification.
Paper: Deep Learning Face Representation by Joint Identification-Verification, Yi Sun et al. (2014). It demonstrates how combining multi-region local part representations with joint identification and verification supervision enhances deep metric embeddings.

Paper: In Defense of the Triplet Loss for Person Re-Identification, Alexander Hermans et al. (2017). It systematically advances triplet metric learning for person re-identification by establishing batch-hard sampling and soft-margin formulations as standard training paradigms.
Paper: Beyond Part Models: Person Retrieval with Refined Part Pooling, Yifan Sun et al. (2017). It refines part-based convolutional representations by replacing rigid multi-channel partitioning with uniform horizontal pooling and adaptive part-feature refinement.
Paper: Learning Discriminative Features with Multiple Granularities for Person Re-Identification, Guanshuo Wang et al. (2018). It builds upon multi-channel global and local body-part architectures by introducing a multi-granularity network with multi-branch stripe divisions and batch-hard loss.
Paper: Harmonious Attention Network for Person Re-identification, Wei Li et al. (2018). It extends joint global and local feature learning in person re-identification by incorporating lightweight multi-granularity harmonious attention mechanisms.
Paper: Re-ranking Person Re-identification with k-Reciprocal Encoding, Zhun Zhong et al. (2017). It provides an unsupervised k-reciprocal encoding post-processing technique to significantly improve rank matching results generated by deep feature embeddings.
Paper: Bag of Tricks and a Strong Baseline for Deep Person Re-Identification, Hao Luo et al. (2019). It demonstrates how systematic optimization tricks and structural refinements like BNNeck can dramatically enhance standard global deep re-identification baselines.
Paper: Deep Learning for Person Re-Identification: A Survey and Outlook, Mang Ye et al. (2020). It provides a comprehensive survey and unified benchmark framework detailing the evolution of deep metric learning, part-based models, and triplet loss designs in person re-identification.

Cover for Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function

Abstract

Person re-identification across cameras remains a very challenging problem, especially when there are no overlapping fields of view between cameras. In this paper, we present a novel multi-channel parts-based convolutional neural network (CNN) model under the triplet framework for person re-identification. Specifically, the proposed CNN model consists of multiple channels to jointly learn both the global full-body and local body-parts features of the input persons. The CNN model is trained by an improved triplet loss function that serves to pull the instances of the same person closer, and at the same time push the instances belonging to different persons farther from each other in the learned feature space. Extensive comparative evaluations demonstrate that our proposed method significantly outperforms many state-of-the-art approaches, including both traditional and deep network-based ones, on the challenging i-LIDS, VIPeR, PRID2011 and CUHK01 datasets.

Topics

convolutional neural networks feature space full body radio channels

Keywords

person re-identification multi-channel parts-based CNN triplet loss function local body-parts features global full-body features metric learning

1. Introduction
2. Related Work
3. The Proposed Person Re-Id Method
3.1. The Overall Framework
3.2. Multi-Channel Parts-based CNN Model
3.3. Improved Triplet Loss Function
3.4. The Training Algorithm
4. Experiments
4.1. Setup
4.2. Experimental Evaluations
4.3. Analysis of different body parts
5. Conclusion
Acknowledgement
References

Knowls

Knowl 1 — Multi-Channel Parts-Based CNN Architecture for Person Re-Identification
model/method
The multi-channel parts-based convolutional neural network (CNN) jointly extracts global full-body representations and local body-part representations for person re-identification. Given an input pedestrian image (resized to $100 \times 250$ pixels and cropped to $80 \times 230$ pixels), the network architecture consists of the following components:
- Global Convolution Layer (G-conv1): A shared initial convolution layer with 32 feature maps, a kernel size of $7 \times 7 \times 3$ , and a stride of 3 pixels, followed by a rectified linear unit (ReLU) activation.
- Full-Body Channel: Takes the full output of G-conv1, applies max pooling with a $3 \times 3$ kernel (B-pool1), feeds into a full-body convolution layer (B-conv2) with 32 feature maps of kernel size $5 \times 5$ (stride 1, ReLU), passes through a second $3 \times 3$ max pooling layer (B-pool2), and connects to a fully connected layer (B-fc) producing a 400-dimensional feature vector.
- Four Body-Part Channels: The feature map output of G-conv1 is vertically divided into four equal non-overlapping horizontal parts $P_i$ for $i \in \{1, 2, 3, 4\}$ . Each part forms the input to an independent body-part channel comprising a convolution layer ( $P_i$ -conv2) with 32 feature maps of kernel size $3 \times 3$ (stride 1, ReLU, no pooling) and a channel-wise fully connected layer ( $P_i$ -fc) producing a 100-dimensional feature vector.
- Network-Wise Fusion (N-fc): The 400-dimensional vector from the full-body channel and the four 100-dimensional vectors from the body-part channels are concatenated into an 800-dimensional vector and passed to a final network-wise fully connected layer that outputs the final 800-dimensional representation $\phi_w(I)$ .
For larger datasets such as CUHK01, a deeper variant is used where each of the five individual channels contains two convolution layers instead of one before its fully connected layer.
Knowl 2 — Improved Triplet Loss Function for Person Re-Identification
equation
Given a triplet of training images $I_i = \langle I_i^o, I_i^+, I_i^- \rangle$ , where $I_i^o$ is an anchor image, $I_i^+$ is a positive match of the same person, and $I_i^-$ is a negative match of a different person, let $\phi_w(I)$ denote the feature embedding produced by the network with parameter set $w$ . The squared $L_2$ -norm distance between two feature representations is:
$d(\phi_w(I_a), \phi_w(I_b)) = \|\phi_w(I_a) - \phi_w(I_b)\|_2^2$
The improved triplet loss enforces both an inter-class margin condition and an intra-class compactness condition across $N$ triplet training examples:
$L(I, w) = \frac{1}{N} \sum_{i=1}^N \left( \max\{d_n(I_i^o, I_i^+, I_i^-, w), \tau_1\} + \beta \max\{d_p(I_i^o, I_i^+, w), \tau_2\} \right)$
where:
- $d_n(I_i^o, I_i^+, I_i^-, w) = d(\phi_w(I_i^o), \phi_w(I_i^+)) - d(\phi_w(I_i^o), \phi_w(I_i^-))$ is the relative inter-class difference, and $\tau_1 < 0$ (set to $\tau_1 = -1$ ) enforces that the intra-class distance is smaller than the inter-class distance by at least $|\tau_1|$ .
- $d_p(I_i^o, I_i^+, w) = d(\phi_w(I_i^o), \phi_w(I_i^+))$ is the intra-class distance, and $\tau_2 > 0$ (set to $\tau_2 = 0.01$ , with $\tau_2 \ll |\tau_1|$ ) restricts positive pairs to lie within a small distance threshold $\tau_2$ to prevent wide intra-class dispersion.
- $\beta \ge 0$ (set to $\beta = 0.002$ ) is a weighting parameter balancing the intra-class compactness penalty relative to the inter-class ranking penalty.
Knowl 3 — Triplet-Based Stochastic Gradient Descent Algorithm for Improved Triplet Loss
algorithm
The parameters $w$ of the multi-channel CNN are trained by stochastic gradient descent minimizing the improved triplet loss function. The gradient with respect to network parameters $w$ is formulated as:
$\frac{\partial L(I, w)}{\partial w} = \frac{1}{N} \sum_{i=1}^N h_1(I_i, w) + \frac{1}{N} \sum_{i=1}^N h_2(I_i, w)$
where the loss-component sub-gradients are:
$h_1(I_i, w) = \begin{cases} \frac{\partial d_n(I_i^o, I_i^+, I_i^-, w)}{\partial w}, & \text{if } d_n(I_i^o, I_i^+, I_i^-, w) > \tau_1 \\ 0, & \text{if } d_n(I_i^o, I_i^+, I_i^-, w) \le \tau_1 \end{cases}$ $h_2(I_i, w) = \begin{cases} \beta \frac{\partial d_p(I_i^o, I_i^+, w)}{\partial w}, & \text{if } d_p(I_i^o, I_i^+, w) > \tau_2 \\ 0, & \text{if } d_p(I_i^o, I_i^+, w) \le \tau_2 \end{cases}$
with the distance gradients calculated from forward activations $\phi_w(\cdot)$ and backward propagation derivatives $\frac{\partial \phi_w(\cdot)}{\partial w}$ :
$\frac{\partial d_p}{\partial w} = 2(\phi_w(I_i^o) - \phi_w(I_i^+)) \frac{\partial \phi_w(I_i^o) - \partial \phi_w(I_i^+)}{\partial w}$ $\frac{\partial d_n}{\partial w} = 2(\phi_w(I_i^o) - \phi_w(I_i^+)) \frac{\partial \phi_w(I_i^o) - \partial \phi_w(I_i^+)}{\partial w} - 2(\phi_w(I_i^o) - \phi_w(I_i^-)) \frac{\partial \phi_w(I_i^o) - \partial \phi_w(I_i^-)}{\partial w}$
The full training procedure is structured as follows:
Input: Training triplet set $\{I_i\}$ , learning rate schedule $\lambda_t$ , maximum iterations $T$ , thresholds $\tau_1, \tau_2$ , weight $\beta$
Output: Learned network parameters $w$
$t \leftarrow 0$
while $t < T$ do
     $t \leftarrow t + 1$
     $\frac{\partial L(I, w)}{\partial w} \leftarrow 0$
    for all training triplet samples $I_i = \langle I_i^o, I_i^+, I_i^- \rangle$ in the mini-batch do
        Calculate $\phi_w(I_i^o)$ , $\phi_w(I_i^+)$ , $\phi_w(I_i^-)$ by forward propagation
        Calculate $\frac{\partial \phi_w(I_i^o)}{\partial w}$ , $\frac{\partial \phi_w(I_i^+)}{\partial w}$ , $\frac{\partial \phi_w(I_i^-)}{\partial w}$ by backpropagation
        Calculate $\frac{\partial d_p}{\partial w}$ and $\frac{\partial d_n}{\partial w}$
        Calculate $h_1(I_i, w)$ and $h_2(I_i, w)$ and accumulate gradient into $\frac{\partial L(I, w)}{\partial w}$
    end for
    Update network parameters: $w^t \leftarrow w^{t-1} - \lambda_t \frac{\partial L(I, w)}{\partial w}$
end while
Knowl 4 — Training Protocols, Triplet Sampling, and Evaluation Setup for Person Re-Identification
experimental setup
- Data Augmentation: All input pedestrian images are normalized to $100 \times 250$ pixels. During training, a bounding box of $80 \times 230$ pixels is cropped around the center with a small random spatial perturbation.
- Weight Initialization: Network weights are initialized from two zero-mean Gaussian distributions with standard deviations of $0.01$ and $0.001$ , respectively; bias terms are initialized to 0.
- Batch Construction and Triplet Sampling: Each mini-batch contains 100 instances constructed by selecting 5 distinct identities and generating 20 triplets per identity per iteration. In each triplet $\langle I^o, I^+, I^- \rangle$ , the positive match $I^+$ is randomly chosen from the same identity class as $I^o$ , and the negative match $I^-$ is randomly chosen from the remaining identity classes. Hyperparameters in the improved triplet loss are set to $\tau_1 = -1$ , $\tau_2 = 0.01$ , and $\beta = 0.002$ .
- Evaluation Protocol: Evaluations follow standard Cumulative Match Characteristic (CMC) curves. For each dataset, approximately half of the identities are randomly sampled for training and the remaining half for testing. In two-camera setups, one image from camera A acts as the query and one from camera B acts as the gallery (one gallery image per person). Matching distances between query and gallery embeddings $\phi_w(I)$ are computed using $L_2$ distance. Top- $k$ accuracy is reported as the average over 10 random train/test splits across four datasets: i-LIDS (119 persons, 479 images), PRID2011 (200 persons appearing in both static cameras A and B), VIPeR (632 persons, 2 camera views), and CUHK01 (971 persons, 4 images each from 2 camera views).
Knowl 5 — Ablation Analysis of Multi-Channel Architecture and Improved Triplet Loss
empirical result
To evaluate the individual contributions of the parts-based multi-channel structure and the intra-class constraint in the improved triplet loss, four model configurations were compared on the i-LIDS, PRID2011, VIPeR, and CUHK01 datasets:
- OursT / Ours3T: Global full-body channel only, trained using the standard triplet loss function without the intra-class constraint ( $\beta = 0$ ).
- OursTC / Ours3TC: Global full-body channel only, trained using the improved triplet loss function ( $\beta = 0.002, \tau_2 = 0.01$ ).
- OursTP / Ours3TP: Complete multi-channel architecture (1 full-body channel + 4 body-part channels), trained using the standard triplet loss function.
- OursTCP / Ours3TCP: Complete multi-channel architecture, trained using the improved triplet loss function.
The comparative results establish two key findings:
1. Effect of Improved Triplet Loss: Adding the intra-class distance constraint improves Rank-1 accuracy by up to $4\%$ across both single-channel and multi-channel architectures (e.g., on i-LIDS, Rank-1 increases from $43.2\%$ for OursT to $47.3\%$ for OursTC, and from $57.2\%$ for OursTP to $60.4\%$ for OursTCP; on VIPeR, Rank-1 increases from $34.3\%$ to $37.2\%$ for single-channel and from $43.8\%$ to $47.8\%$ for multi-channel).
2. Effect of Body-Part Channels: Jointly learning body-part features alongside the full body yields an accuracy boost of up to $13.1\%$ over the full-body channel alone (e.g., on i-LIDS, Rank-1 increases from $47.3\%$ for OursTC to $60.4\%$ for OursTCP; on VIPeR, Rank-1 increases from $37.2\%$ for OursTC to $47.8\%$ for OursTCP).

Knowl 6 — Cumulative Match Characteristic Performance on Standard Re-Identification Datasets

data/table

The proposed full model (OursTCP for i-LIDS, PRID2011, and VIPeR; Ours3TCP for CUHK01) was evaluated against representative metric learning, salience-based, and deep CNN methods using the Cumulative Match Characteristic (CMC) ranking metric across top-1 to top-30/100 ranks under standard 50/50 train/test splits averaged over 10 trials.

i-LIDS Dataset	Top1	Top5	Top10	Top15	Top20	Top30
Adaboost	29.6	55.2	68.1	77.0	82.4	92.1
LMNN	28.0	53.8	66.1	75.5	82.3	91.0
ITML	29.0	54.0	70.5	81.0	86.7	95.0
MCC	31.3	59.3	75.6	84.0	88.3	95.0
PRDC	37.8	63.7	75.1	82.8	88.4	95.0
Sakrapee (Metric Ensemble)	50.3	–	–	–	–	–
Ding (Deep Relative Comparison)	52.1	68.2	78.0	83.6	88.8	95.0
OursTCP	60.4	82.7	90.7	96.4	97.8	99.3
PRID2011 Dataset	Top1	Top10	Top20	Top50	Top100
KISSME	15.0	39.0	52.0	68.0	80.0
EIML	16.0	39.0	51.0	68.0	81.0
LMNN	10.0	30.0	42.0	59.0	73.0
ITML	12.0	36.0	47.0	64.0	79.0
Maha	16.0	41.0	51.0	64.0	76.0
DeepM	17.9	45.9	55.4	71.4	–
Sakrapee	17.9	–	–	–	–
OursTCP	22.0	47.0	57.0	76.0	83.0
VIPeR Dataset	Top1	Top5	Top10	Top15	Top20	Top30
MtMCML	28.8	59.3	75.8	83.4	88.5	93.5
SDALF	19.9	38.4	49.4	58.5	66.0	74.4
eSDC	26.3	46.4	58.6	66.6	72.8	80.5
SalMatch	30.2	52.3	66.0	73.4	79.2	86.0
Ding	40.5	60.8	70.4	78.3	84.4	90.9
mFilter+LADF	43.4	–	–	–	–	–
Sakrapee	45.9	–	–	–	–	–
OursTCP	47.8	74.7	84.8	89.2	91.1	94.3
CUHK01 Dataset	Top1	Top5	Top10	Top15	Top20	Top30
mFilter	34.3	55.0	65.3	70.5	–	–
SalMatch	28.5	46.3	57.2	64.1	–	–
FPNN	27.9	–	–	–	–	–
Ejaz	47.5	–	–	–	–	–
Sakrapee	53.4	76.4	84.4	–	90.5	–
Ours3TCP	53.7	84.3	91.0	93.3	96.3	98.3

The proposed method outperforms all baseline traditional and deep network models across all ranking measures on the four benchmark datasets, outperforming the previous state-of-the-art ensemble method by $10.1\%$ Rank-1 accuracy on i-LIDS ( $60.4\%$ vs. $50.3\%$ ), $4.1\%$ on PRID2011 ( $22.0\%$ vs. $17.9\%$ ), and $1.9\%$ on VIPeR ( $47.8\%$ vs. $45.9\%$ ).

Knowl 7 — Sensitivity of Re-Identification Accuracy to the Intra-Class Loss Weight Hyperparameter

data/table

The hyperparameter $\beta$ in the improved triplet loss balances the relative importance of the intra-class compactness constraint and the inter-class separation constraint. A cross-validation evaluation on the VIPeR dataset demonstrates how identification accuracy (%) varies across different values of $\beta$ :

$\beta$	Top1	Top5	Top10	Top15	Top20	Top30
0	43.8	69.5	79.7	81.0	85.4	90.2
0.001	45.9	73.4	81.9	87.0	93.0	95.6
0.002	47.8	74.7	84.8	89.2	91.1	94.3
0.003	45.6	75.3	85.4	87.6	90.5	94.6
0.004	43.7	73.1	81.5	87.9	91.1	93.4

Setting $\beta = 0$ corresponds to the conventional triplet loss without an intra-class constraint. Re-identification performance peaks within the interval $\beta \in [0.001, 0.003]$ , with $\beta = 0.002$ providing the highest Rank-1 identification rate ( $47.8\%$ ). Increasing $\beta$ beyond $0.003$ causes performance to decline, as an excessive intra-class penalty impairs the model's ability to maintain sufficient inter-class margin separation.

Knowl 8 — Spatial Body-Part Contribution to Re-Identification Accuracy
empirical result
To evaluate how different spatial body regions contribute to re-identification accuracy, four distinct two-channel network models were trained on the VIPeR dataset. Each model combines the full-body channel with exactly one of the four body-part channels ( $P_1, P_2, P_3, P_4$ , partitioned from top to bottom of the pedestrian):
- Ours-Part1 (Face and Shoulders): Produces the largest performance improvement over the full-body baseline (OursT).
- Ours-Part2 (Upper Torso / Chest): Provides the second largest performance gain.
- Ours-Part3 (Lower Torso / Hips): Yields moderate performance gain.
- Ours-Part4 (Legs and Feet): Provides the lowest performance improvement among all four parts.
The marginal benefit of local part features decreases monotonically from the top to the bottom of the body. This is because the head and upper torso maintain relatively stable appearance and shape across camera viewpoints, whereas the legs and feet undergo large deformations, motion blur, and pose variations during walking, yielding less reliable features for person re-identification. Combining all four parts with the full-body channel (OursTP) surpasses every single-part variant.

Coverage note — All substantial contributed materials—including the multi-channel network architecture, the improved triplet loss formulation, the SGD training algorithm, the experimental protocol and parameters, ablation analyses, benchmark comparison tables, hyperparameter sensitivity analysis, and the body-part contribution study—have been included. The qualitative visualization of convolution activations was omitted as it merely visualizes the learned features without adding standalone technical content.

References

1.E. Ahmed, M. Jones, and T. K. Marks. An improved deep learning architecture for person re-identification. CVPR, 5:25, 2015.
2.S. Bak, E. Corvee, F. Bremond, and M. Thonnat. Person re-identification using spatial covariance regions of human body parts. In Advanced Video and Signal Based Surveillance (AVSS), 2010 Seventh IEEE International Conference on, pages 435–440, 2010.
3.X. Chang, Y. Yang, E. P. Xing, and Y.-L. Yu. Complex event detection using semantic saliency and nearly-isotonic svm. In International Conference on Machine Learning (ICML), 2015.
4.D. S. Cheng, M. Cristani, M. Stoppa, L. Bazzani, and V. Murino. Custom pictorial structures for re-identification. In BMVC, volume 1, page 6, 2011.
5.J. V. Davis, B. Kulis, P. Jain, S. Sra, and I. S. Dhillon. Information-theoretic metric learning. In ICML, pages 209–216, 2007.
6.S. Ding, L. Lin, G. Wang, and H. Chao. Deep feature learning with relative distance comparison for person re-identification. Pattern Recognition, 2015.
7.P. Dollár, Z. Tu, H. Tao, and S. Belongie. Feature mining for image classification. In CVPR, pages 1–8, 2007.
8.M. Farenzena, L. Bazzani, A. Perina, V. Murino, and M. Cristani. Person re-identification by symmetry-driven accumulation of local features. In CVPR, pages 2360–2367, 2010.
9.P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan. Object detection with discriminatively trained part-based models. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 32(9):1627–1645, 2010.
10.N. Gheissari, T. B. Sebastian, and R. Hartley. Person reidentification using spatiotemporal appearance. In CVPR, volume 2, pages 1528–1535, 2006.
11.R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, pages 580–587, 2014.
12.A. Globerson and S. T. Roweis. Metric learning by collapsing classes. In NIPS, pages 451–458, 2005.
13.D. Gray, S. Brennan, and H. Tao. Evaluating appearance models for recognition, reacquisition, and tracking. In Proc. IEEE International Workshop on Performance Evaluation for Tracking and Surveillance (PETS), volume 3, 2007.
14.D. Gray and H. Tao. Viewpoint invariant pedestrian recognition with an ensemble of localized features. In ECCV, pages 262–275. 2008.
15.M. Guillaumin, J. Verbeek, and C. Schmid. Is that you? metric learning approaches for face identification. In CVPR, pages 498–505, 2009.
16.J. Han, D. Zhang, S. Wen, L. Guo, T. Liu, and X. Li. Two-stage learning to predict human eye fixations via sdaes. 2015.
17.M. Hirzer, C. Beleznai, P. M. Roth, and H. Bischof. Person re-identification by descriptive and discriminative classification. In Image Analysis, pages 91–102. 2011.
18.M. Hirzer, P. M. Roth, and H. Bischof. Person re-identification by efficient impostor-based metric learning. In Advanced Video and Signal-Based Surveillance (AVSS), 2012 IEEE Ninth International Conference on, pages 203–208, 2012.
19.W. Hu, M. Hu, X. Zhou, T. Tan, J. Lou, and S. Maybank. Principal axis-based correspondence between multiple cameras for people tracking. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 28(4):663–671, 2006.
20.S. Khamis, C.-H. Kuo, V. K. Singh, V. D. Shet, and L. S. Davis. Joint learning for attribute-consistent person re-identification. In ECCV, pages 134–146, 2014.
21.M. Koestinger, M. Hirzer, P. Wohlhart, P. M. Roth, and H. Bischof. Large scale metric learning from equivalence constraints. In CVPR, pages 2288–2295, 2012.
22.A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, pages 1097–1105, 2012.
23.W. Li and X. Wang. Locally aligned feature transforms across views. In CVPR, pages 3594–3601, 2013.
24.W. Li, R. Zhao, and X. Wang. Human reidentification with transferred metric learning. In ACCV, pages 31–44, 2012.
25.W. Li, R. Zhao, T. Xiao, and X. Wang. Deepreid: Deep filter pairing neural network for person re-identification. In CVPR, pages 152–159, 2014.
26.Z. Li, S. Chang, F. Liang, T. S. Huang, L. Cao, and J. R. Smith. Learning locally-adaptive decision functions for person verification. In CVPR, pages 3610–3617, 2013.
27.C. Liu, S. Gong, C. C. Loy, and X. Lin. Person re-identification: What features are important? In ECCV, pages 391–401, 2012.
28.B. Ma, Y. Su, and F. Jurie. Bicov: a novel image representation for person re-identification and face verification. In BMVC, pages 11–pages, 2012.
29.L. Ma, X. Yang, and D. Tao. Person re-identification over camera networks using multi-task distance metric learning. Image Processing, IEEE Transactions on, 23(8):3656–3670, 2014.
30.A. Mignon and F. Jurie. Pcca: A new approach for distance learning from sparse pairwise constraints. In CVPR, pages 2666–2672, 2012.
31.S. Paisitkriangkrai, C. Shen, and A. v. d. Hengel. Learning to rank in person re-identification with metric ensembles. arXiv preprint arXiv:1503.01543, 2015.
32.U. Park, A. K. Jain, I. Kitahara, K. Kogure, and N. Hagita. Vise: Visual search engine using multiple networked cameras. In ICPR, volume 3, pages 1204–1207, 2006.
33.P. M. Roth, M. Hirzer, M. Köstinger, C. Beleznai, and H. Bischof. Mahalanobis distance learning for person re-identification. In Person Re-Identification, pages 247–267. 2014.
34.F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. arXiv preprint arXiv:1503.03832, 2015.
35.W. R. Schwartz and L. S. Davis. Learning discriminative appearance-based models using partial least squares. In Computer Graphics and Image Processing (SIBGRAPI), 2009 XXII Brazilian Symposium on, pages 322–329, 2009.
36.Y. Sun, Y. Chen, X. Wang, and X. Tang. Deep learning face representation by joint identification-verification. In NIPS, pages 1988–1996, 2014.
37.A. Toshev and C. Szegedy. Deeppose: Human pose estimation via deep neural networks. In CVPR, pages 1653–1660, 2014.
38.UK. Home office i-lids multiple camera tracking scenario definition. In 2008.
39.J. Wang, Y. Song, T. Leung, C. Rosenberg, J. Wang, J. Philbin, B. Chen, and Y. Wu. Learning fine-grained image similarity with deep ranking. In CVPR, pages 1386–1393, 2014.
40.X. Wang, G. Doretto, T. Sebastian, J. Rittscher, and P. Tu. Shape and appearance context modeling. In ICCV, pages 1–8, 2007.
41.K. Q. Weinberger, J. Blitzer, and L. K. Saul. Distance metric learning for large margin nearest neighbor classification. In NIPS, pages 1473–1480, 2005.
42.E. P. Xing, M. I. Jordan, S. Russell, and A. Y. Ng. Distance metric learning with application to clustering with side-information. In NIPS, pages 505–512, 2002.
43.F. Xiong, M. Gou, O. Camps, and M. Sznaier. Person re-identification using kernel-based metric learning methods. In ECCV, pages 1–16. 2014.
44.Y. Yang, J. Yang, J. Yan, S. Liao, D. Yi, and S. Z. Li. Salient color names for person re-identification. In ECCV, pages 536–551. 2014.
45.D. Yi, Z. Lei, S. Liao, and S. Z. Li. Deep metric learning for person re-identification. In ICPR, pages 34–39, 2014.
46.R. Zhao, W. Ouyang, and X. Wang. Person re-identification by salience matching. In ICCV, pages 2528–2535, 2013.
47.R. Zhao, W. Ouyang, and X. Wang. Unsupervised salience learning for person re-identification. In CVPR, pages 3586–3593, 2013.
48.R. Zhao, W. Ouyang, and X. Wang. Learning mid-level filters for person re-identification. In CVPR, pages 144–151, 2014.
49.W.-S. Zheng, S. Gong, and T. Xiang. Person re-identification by probabilistic relative distance comparison. In CVPR, pages 649–656, 2011.

Citation

MLA

Cheng, D., et al. “Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function”. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 1335–44, https://doi.org/10.1109/CVPR.2016.149.

APA

Cheng, D., Gong, Y., Zhou, S., Wang, J., & Zheng, N. (2016). Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1335–1344. https://doi.org/10.1109/CVPR.2016.149

Chicago

Cheng, D., Y. Gong, S. Zhou, J. Wang, and N. Zheng. 2016. “Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function”. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1335–44. https://doi.org/10.1109/CVPR.2016.149.

Harvard

Cheng, D. et al. (2016) “Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function”, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 1335–1344. Available at: https://doi.org/10.1109/CVPR.2016.149.

Vancouver

1. Cheng D, Gong Y, Zhou S, Wang J, Zheng N (2016) Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp 1335–1344

BibTeX

@inproceedings{Cheng_2016, title={Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function}, url={http://dx.doi.org/10.1109/CVPR.2016.149}, DOI={10.1109/cvpr.2016.149}, booktitle={2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}, publisher={IEEE}, author={Cheng, De and Gong, Yihong and Zhou, Sanping and Wang, Jinjun and Zheng, Nanning}, year={2016}, month=June, pages={1335–1344} }

Metadata:Crossref

Access the Paper

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

Open PDF

License: IEEE

Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Multi-Channel Parts-Based CNN Architecture for Person Re-Identification

Knowl 2 — Improved Triplet Loss Function for Person Re-Identification

Knowl 3 — Triplet-Based Stochastic Gradient Descent Algorithm for Improved Triplet Loss

Knowl 4 — Training Protocols, Triplet Sampling, and Evaluation Setup for Person Re-Identification

Knowl 5 — Ablation Analysis of Multi-Channel Architecture and Improved Triplet Loss