[論文レビュー] Triplet Contrastive Representation Learning for Unsupervised Vehicle Re-identification

Fei Shen, Xiaoyu Du+3 more|arXiv (Cornell University)|Jan 23, 2023

Advanced Neural Network Applications被引用数 10

ひとこと要約

本論文は、三つのメモリーバンク（part、cluster、global）と三つの損失（PCL、HCL、WRCCL）を用いたトリプレット対照学習フレームワーク、TCRLを提案し、部位特徴とグローバル特徴を結びつけて教師なしの車両再識別を行い、VeRi776、VehicleID、VERI-Wildで追加データなしに最先端の結果を達成します。

ABSTRACT

Part feature learning is critical for fine-grained semantic understanding in vehicle re-identification. However, existing approaches directly model part features and global features, which can easily lead to serious gradient vanishing issues due to their unequal feature information and unreliable pseudo-labels for unsupervised vehicle re-identification. To address this problem, in this paper, we propose a simple Triplet Contrastive Representation Learning (TCRL) framework which leverages cluster features to bridge the part features and global features for unsupervised vehicle re-identification. Specifically, TCRL devises three memory banks to store the instance/cluster features and proposes a Proxy Contrastive Loss (PCL) to make contrastive learning between adjacent memory banks, thus presenting the associations between the part and global features as a transition of the part-cluster and cluster-global associations. Since the cluster memory bank copes with all the vehicle features, it can summarize them into a discriminative feature representation. To deeply exploit the instance/cluster information, TCRL proposes two additional loss functions. For the instance-level feature, a Hybrid Contrastive Loss (HCL) re-defines the sample correlations by approaching the positive instance features and pushing the all negative instance features away. For the cluster-level feature, a Weighted Regularization Cluster Contrastive Loss (WRCCL) refines the pseudo labels by penalizing the mislabeled images according to the instance similarity. Extensive experiments show that TCRL outperforms many state-of-the-art unsupervised vehicle re-identification approaches.

研究の動機と目的

• 教師なし再識別において部位とグローバル車両特徴を学習する際の勾配消失および信頼性の低い pseudo-label を解決する。
• クラスタベースのプロキシを通じて部位レベルとグローバルレベルの表現を橋渡しし、効果的なクロス-feature 学習を可能にする。
• インスタンスレベルおよびクラスタレベルの情報を活用するために、メモリーバンクと新規搾損失を活用して識別性の高い車両表現を構築する。
• 追加のラベル付きデータなしで大規模車两再識別データセットで最先端の性能を示す。

提案手法

• Three memory banks store part features (M^P), cluster representatives (M^C), and global features (M^G).
• Proxy Contrastive Loss (PCL) models associations between adjacent memory banks (part–cluster, cluster–global) using KL-divergence and Euclidean distance.
• Hybrid Contrastive Loss (HCL) integrates all negative instance samples to pull together positives at the instance level.
• Weighted Regularization Cluster Contrastive Loss (WRCCL) weights cluster-level pseudo-label correlations by image similarity to suppress mislabeled samples.
• Cluster memory bank summarizes features to provide a discriminative representation and guides pseudo-label updates via DBSCAN-based clustering.
• Overall loss L_Total combines PCL, HCL, and WRCCL with balancing coefficients.
• Training uses masked and original image pairs, momentum-updated memories, and a ResNet-50 backbone.
• Evaluation is conducted with cosine similarity on learned representations in an unsupervised setting across VeRi776, VehicleID, and VERI-Wild.

実験結果

リサーチクエスチョン

• RQ1 Can a triplet contrastive framework effectively connect part-level and global-level vehicle features in an unsupervised setting?
• RQ2 Do memory-bank-based proxy and carefully designed loss functions improve unsupervised vehicle re-identification performance?
• RQ3 How do Part, Cluster, and Global memories interact to produce robust, discriminative representations for fine-grained vehicle re-id?

主な発見

• TCRL achieves state-of-the-art unsupervised performance on VeRi776 with mAP 42.68% and Rank-1 87.26%.
• On VehicleID, TCRL reaches mAP 66.29% and Rank-1 60.36% on Test800, with strong results across larger Test1600 and Test2400 subsets (values in the paper).
• On VERI-Wild, TCRL attains best results across Test3000, Test5000, and Test10000 subsets (e.g., mAP 66.29% and Rank-1 60.36% on Test3000).
• Ablation studies show directly modeling part+global features collapses without the cluster proxy, while incorporating part features and the proposed losses (PCL, HCL, WRCCL) consistently improves performance across datasets.
• The combined TCRL objective (L_Total with L_PCL, L_HCL, and L_WRCCL) outperforms baselines and individual losses, demonstrating the benefit of triplet contrastive learning with three memory banks.