How Multi-Level Similarity‑Aware CNN Boosts Person Re‑Identification Accuracy

Abstract

Person re‑identification (person re‑ID) aims to match images of the same individual captured by non‑overlapping cameras. The paper proposes a novel deep siamese architecture called Multi‑Level Similarity‑Aware CNN (MSP‑CNN) that applies different similarity constraints to low‑level and high‑level feature maps during training, enabling discriminative feature learning and significantly improving re‑ID performance.

The framework also integrates classification constraints, allowing a unified multi‑task network that does not require paired inputs at test time; features can be extracted offline to build an index for large‑scale deployment.

1 Introduction

Person re‑ID is challenging because different identities may look similar while the same identity can appear under varying illumination, viewpoints, and occlusions. Existing CNN‑based methods have shown strong performance, but most ignore the distinct characteristics of feature maps at different depths.

MSP‑CNN introduces a siamese model that processes image pairs through shared CNN parameters, using small convolution filters and Inception modules. It applies low‑level similarity constraints on the Pool1 layer and high‑level constraints on the FC7 layer, leveraging both local and global cues.

2 Proposed Method

2.1 Overview

The base CNN is designed with three CONV modules, six Inception modules, and one fully‑connected module. Input images are resized to 160×64 and randomly cropped to 144×56 for data augmentation.

Low‑level similarity is enforced on the Pool1 feature map using normalized cross‑correlation, while high‑level similarity is enforced on the FC7 features using Euclidean distance after L2 normalization.

2.2 Multi‑Level Similarity Awareness

Low‑level constraints focus on discriminative local patches (e.g., a red backpack) that appear across images of the same person. High‑level constraints encourage global feature similarity.

2.3 Multi‑Task Architecture

Both similarity and classification (softmax) losses are combined in a unified network. During training, the network first learns with softmax and Euclidean losses, then adds the low‑level cross‑correlation loss for a few epochs to avoid over‑fitting.

At test time, each gallery image’s feature is extracted once; queries are ranked by Euclidean distance, and indexing techniques (e.g., inverted index or hashing) can be applied for efficiency.

3 Experiments

3.1 Datasets and Protocols

Evaluations are conducted on CUHK03, Market‑1501, and CUHK01 using Cumulative Matching Characteristic (CMC) top‑k accuracy and mean Average Precision (mAP) for Market‑1501.

3.2 Implementation Details

The model is implemented in Caffe and trained on an NVIDIA M40 GPU. The network uses small 3×3 filters, batch normalization, ReLU, and Inception‑v3‑style modules.

3.3 Training Strategy

A sampling strategy maintains a 2:1 ratio of negative to positive pairs. Data augmentation follows AlexNet techniques.

4 Results and Discussion

4.1 Comparison with State‑of‑the‑Art

MSP‑CNN outperforms existing methods on all three datasets, achieving higher rank‑1, rank‑5, and mAP scores.

4.2 Ablation Study

Experiments on CUHK03 with annotated bounding boxes show the contribution of each component: baseline classification, high‑level similarity, low‑level similarity, and their combination.

5 Conclusion and Future Work

The authors plan to explore suitable optimization objectives for middle‑level layers and to leverage more feature maps for further performance gains.

References

Xiao et al., “Learning deep feature representations with domain guided dropout for person re‑identification,” CVPR 2016.

Subramaniam et al., “Deep Neural Networks with Inexact Matching for Person Re‑Identification,” NIPS 2016.

Krizhevsky et al., “ImageNet classification with deep convolutional neural networks,” NIPS 2012.