[論文レビュー] Multi-task Mid-level Feature Alignment Network for Unsupervised Cross-Dataset Person Re-Identification

Shan Lin, Haoliang Li+2 more|arXiv (Cornell University)|Jul 3, 2018

Video Surveillance and Tracking Methods被引用数 159

ひとこと要約

本論文は MMFA を提案する。エンドツーエンドの無監視クロスデータセット人物再識別フレームワークで、MMDベースの正則化を介してドメイン間の中間レベル特徴を整列させつつ、同時に識別と属性監督を jointly 学習する。

ABSTRACT

Most existing person re-identification (Re-ID) approaches follow a supervised learning framework, in which a large number of labelled matching pairs are required for training. Such a setting severely limits their scalability in real-world applications where no labelled samples are available during the training phase. To overcome this limitation, we develop a novel unsupervised Multi-task Mid-level Feature Alignment (MMFA) network for the unsupervised cross-dataset person re-identification task. Under the assumption that the source and target datasets share the same set of mid-level semantic attributes, our proposed model can be jointly optimised under the person's identity classification and the attribute learning task with a cross-dataset mid-level feature alignment regularisation term. In this way, the learned feature representation can be better generalised from one dataset to another which further improve the person re-identification accuracy. Experimental results on four benchmark datasets demonstrate that our proposed method outperforms the state-of-the-art baselines.

研究の動機と目的

• Motivate scalable person Re-ID without target-domain labels by leveraging shared mid-level attributes across datasets.
• Learn discriminative features through joint identity and attribute supervision on a labeled source domain.
• Align source and target mid-level feature distributions to improve cross-dataset generalization.
• Introduce a single-stage end-to-end training framework that combines feature learning and domain adaptation.

提案手法

• Use a ResNet-50 backbone with a global max-pooling layer and multiple task-specific heads.
• Train with identity classification loss and multiple attribute classification losses.
• Compute two domain-adaptation regularizers based on MMD: one on attribute features (AAL) and one on mid-level deep features (MDAL).
• Combine losses into a unified objective L_all = L_id + λ1 L_attr + λ2 L_AAL + λ3 L_MDAL.
• Employ a mixture of RBF kernels for MMD and train in a single end-to-end SGD process on mini-batches containing both labeled source and unlabeled target samples.

実験結果

リサーチクエスチョン

• RQ1 Can unsupervised cross-dataset Re-ID be achieved by aligning mid-level feature distributions across domains?
• RQ2 Does jointly learning identity and mid-level attributes improve cross-dataset generalization?
• RQ3 How effective are MMD-based regularizers on attributes and mid-level features for domain adaptation in Re-ID?
• RQ4 How does MMFA perform compared to state-of-the-art unsupervised cross-dataset Re-ID methods on standard benchmarks?

主な発見

• MMFA outperforms many state-of-the-art unsupervised methods on VIPeR, PRID, Market1501 and DukeMTMC-reID benchmarks.
• Attribute and identity joint training provides better generalisation than using either alone (Attribute+ID yields improvements over Attribute Only).
• Mid-level feature alignment (AAL and MDAL) significantly boosts cross-domain performance, with substantial gains after alignment compared to non-adapted features.
• On Market1501 and DukeMTMC-reID, MMFA achieves notable Rank-1 and mAP improvements over several baselines (e.g., MMFA Duke: Rank-1 56.7, mAP 27.4 in cross-dataset setup; MMFA Market: Rank-1 45.3, mAP 24.7).
• The approach requires only one end-to-end training session (about 25 epochs) to reach competitive performance.