Source-linked AI summary
Person Re-Identification by Deep Joint Learning of Multi-Loss Classification
Wei Li, Xiatian Zhu, Shaogang Gong
TL;DR
Existing person re-identification methods often rely on either local or global features, leaving their complementary benefits underused. JLML jointly learns both through interacting branches and separate classification objectives under shared identity supervision. Evaluations report advantages over state-of-the-art methods across five benchmarks.
Problem
Existing re-identification methods often rely on either local or global feature representations, despite the potential complementary benefits of learning both jointly.
Method
JLML is a two-branch CNN that jointly learns interacting local and global feature selections with separate classification losses and joint sparsity regularisation under shared identity labels.
Results
Extensive comparative evaluations report JLML superiority over a wide range of state-of-the-art re-identification models on VIPeR, GRID, CUHK01, CUHK03, and Market-1501.
Takeaways & Limitations
Jointly learned local and global representations can exploit complementary fine-grained saliency and holistic context for person re-identification.
Takeaways & Limitations
The implementation found no clear benefit from increasing the number of shared convolutional layers, and the base network choice is treated as independent of JLML design.
Abstract
from arXiv · showhide
Existing person re-identification (re-id) methods rely mostly on either localised or global feature representation alone. This ignores their joint benefit and mutual complementary effects. In this work, we show the advantages of jointly learning local and global features in a Convolutional Neural Network (CNN) by aiming to discover correlated local and global features in different context. Specifically, we formulate a method for joint learning of local and global feature selection losses designed to optimise person re-id when using only generic matching metrics such as the L2 distance. We design a novel CNN architecture for Jointly Learning Multi-Loss (JLML) of local and global discriminative feature optimisation subject concurrently to the same re-id labelled information. Extensive comparative evaluations demonstrate the advantages of this new JLML model for person re-id over a wide range of state-of-the-art re-id methods on five benchmarks (VIPeR, GRID, CUHK01, CUHK03, Market-1501).
1 Introduction
Person re-identification must handle severe cross-view appearance changes, yet existing methods often learn either local or global representations. The paper proposes JLML, a CNN that jointly learns complementary local and global feature selections under shared identity supervision.
- Person re-identification matches identity classes across non-overlapping camera views despite changes in pose, illumination, occlusion, and background clutter.
- Existing approaches commonly emphasize feature representation, matching metrics, or their combination within deep learning frameworks.
- Local and global feature learning are jointly motivated because local saliency and global context provide correlated, complementary information under the same identity-label constraint.
- The model jointly optimises local and global feature selections for re-identification using generic matching metrics such as L2 distance.
- JLML uses a two-branch CNN with local and global feature-selection branches that interact while retaining separate classification losses.
- Structured sparsity-based feature selection is introduced to improve robustness to noise and covariance between local and global representations.
2 Related Works
JLML extends prior local-feature and global-feature approaches by jointly learning both representation levels. Unlike related methods using supervised metric learning or separately extracted hand-crafted features, it learns discriminative features with generic matching metrics.
- Earlier saliency-based methods model localised part importance but consider only patch appearance statistics within individual regions.
- SCS and MCP consider local and global representation levels, but JLML differs by jointly learning discriminative features rather than focusing on supervised metric learning.
- JLML uses generic matching metrics, whereas SCS performs supervised metric learning and extracts hand-crafted local and global features separately.
- HER optimises re-identification embeddings with regression from a single identity label, but relies on predefined features and cannot jointly learn global and local representations.
3 Model Design
JLML uses a two-branch CNN to learn local and global person features jointly under shared identity supervision, while separate branch losses preserve complementary discriminative capabilities. Structured sparsity selects useful features, and the resulting concatenated representation is matched with a generic distance such as L2.
- Network Construction: JLML-ResNet39 combines m local streams for body regions with a global branch processing the entire person image.The architecture uses residual CNN building blocks and is customised as a 39-layer network.
- Joint Learning Multi-Loss: The branches interact while retaining separate objective losses, enabling independent local and global learning under the same identity-label constraint.This branch-wise formulation is called the MultiLoss design.
- Feature Selection: Structured sparsity regularisation selectively suppresses irrelevant features and promotes discriminative representations in both global and local contexts.The global branch uses group LASSO, while the local branch uses an exclusive group LASSO formulation.
- Loss Function: Cross-entropy classification losses train both branches using person identity labels, with sparsity weights balancing classification and feature-selection regularisation.The global and local coefficients are empirically set to λlocal = λglobal = 5×10−4.
- Joint Optimisation: SGD synchronises branch updates under shared identity information, reducing the risk that local and global representations diverge during learning.The model is trained batch-wise with the two branches jointly optimised.
- Re-identification Matching: The learned 1,024-D representation concatenates 512-D local and global vectors, which are separately normalised and compared across cameras using a generic metric such as L2 distance.Gallery images are ranked by ascending cross-camera distance to the probe.
4 Experiments
Experiments evaluate JLML across five re-id benchmarks and analyze its efficiency, feature complementarity, branch design, matching metrics, decomposition, and extraction cost. JLML consistently outperforms competing methods while retaining relatively low model complexity and fast feature extraction.
- Datasets and protocol: JLML was evaluated on VIPeR, GRID, CUHK01, CUHK03, and Market-1501 under varied population sizes and challenging viewing conditions.The evaluation uses standard supervised re-id protocols with dataset-specific training and test splits.
- Comparative evaluation: JLML outperformed existing methods on CUHK03, Market-1501, CUHK01, and GRID, including a 12.8% Rank-1 gain over GOG on GRID.On Market-1501, JLML gained 19.3% Rank-1 in single-query and 13.7% in multi-query over S-CNN; CUHK01 results were also superior across both splits and shot settings.
- Efficiency and deployment: JLML-ResNet39 has 7.2 million parameters and 1.54×10^9 FLOPs, both second-smallest among the compared CNN architectures despite using five streams.Feature extraction averaged 2.75 milliseconds per image, or 364 images per second, on an Nvidia Pascal P100 GPU.
- Component analysis: The joint local-global representation improved Market-1501 performance over either feature alone, adding up to 6.2% Rank-1 and 7.7% mAP in single-query evaluation.The corresponding multi-query gains were 3.3% Rank-1 and 6.1% mAP.
- Component analysis: Independent branch losses improved Market-1501 performance over a merged UniLoss design by 9.0% Rank-1 and 13.3% mAP in single-query evaluation.The authors attribute the degradation under UniLoss to a single loss affecting branches whose discriminative features may diverge across local and global contexts.
- Design analysis: Selective Feature Learning added 1.7% Rank-1 and 1.7% mAP in single-query evaluation, while L1 and L2 distances produced similar competitive accuracies.Four body parts were the optimal local decomposition choice.
5 Conclusion
The paper presents JLML-ResNet39, a CNN that jointly learns global and local discriminative features through multi-loss classification. It also uses structured sparsity for feature selection and validates the model across five person re-identification benchmarks.
- JLML-ResNet39 jointly learns global and local discriminative feature representations within a unified multi-loss classification framework.
- The architecture exploits correlated complementary information from different visual contexts rather than relying on global or local features alone.
- A structured sparsity-based feature-selection mechanism reduces feature redundancy and improves joint feature selection.
- Evaluations on five person re-identification benchmarks compare JLML with a wide range of state-of-the-art methods.
- The study includes component evaluations and model-performance analyses covering manually labelled and more challenging auto-detected person images.