[Paper Review] Unsupervised Domain Adaptive Re-Identification: Theory and Practice

Liangchen Song, Cheng Wang+5 more|arXiv (Cornell University)|Jul 30, 2018

Domain Adaptation and Few-Shot LearningComputer Science16 references66 citations

TL;DR

The paper provides a theory-guided framework for unsupervised domain adaptive re-ID, introducing feature-space assumptions and a self-training scheme using clustering to produce pseudo-labels, validated on person and vehicle re-ID tasks.

ABSTRACT

We study the problem of unsupervised domain adaptive re-identification (re-ID) which is an active topic in computer vision but lacks a theoretical foundation. We first extend existing unsupervised domain adaptive classification theories to re-ID tasks. Concretely, we introduce some assumptions on the extracted feature space and then derive several loss functions guided by these assumptions. To optimize them, a novel self-training scheme for unsupervised domain adaptive re-ID tasks is proposed. It iteratively makes guesses for unlabeled target data based on an encoder and trains the encoder based on the guessed labels. Extensive experiments on unsupervised domain adaptive person re-ID and vehicle re-ID tasks with comparisons to the state-of-the-arts confirm the effectiveness of the proposed theories and self-training framework. Our code is available at \url{https://github.com/LcDog/DomainAdaptiveReID}.

Motivation & Objective

• Extend unsupervised domain adaptation theory to re-ID tasks by adapting input-space assumptions to pairwise feature space.
• Introduce three feature-space assumptions (covariate shift, Separately Probabilistic Lipschitzness, weight ratio) and derive corresponding losses.
• Propose a self-training framework to iteratively refine the encoder using pseudo-labels.
• Design clustering- and distance-based strategies to select high-confidence target samples for training.
• Demonstrate effectiveness on large-scale person and vehicle re-ID datasets.

Proposed method

• Adapt covariate shift to pairwise re-ID labels.
• Define Separately Probabilistic Lipschitzness (SPL) for feature pairs and derive intra-/inter-cluster losses (L_intra, L_inter).
• Incorporate a weight-ratio based loss (L_WR) to align source/target feature distributions.
• Use a distance metric combining Jaccard-based similarity (d_J) and a target-to-source closeness term (d_W) to form a joint matrix M for clustering.
• Perform DBSCAN clustering on the distance matrix to select pseudo-labeled target samples.
• Train with triplet loss on selected pseudo-labeled data and iterate to update the encoder.

Experimental results

Research questions

• RQ1 Can three feature-space assumptions (covariate shift, SPL, weight ratio) yield learnability guarantees for unsupervised domain adaptive re-ID?
• RQ2 How can an encoder be trained to satisfy SPL and exploit weight ratio in an iterative, self-training framework?
• RQ3 Does a clustering-based pseudo-labeling strategy improve cross-domain re-ID performance over direct transfer and naive self-training?
• RQ4 How effective is the proposed distance metric and clustering approach (DBSCAN with k-reciprocal encoding) on person and vehicle re-ID across domains?

Key findings

• Theoretical guarantees for unsupervised domain adaptive re-ID under three feature-space assumptions.
• A self-training framework that iteratively refines the encoder using pseudo-labels derived from clustering.
• A distance-metric scheme combining d_J and d_W that improves clusterability and pseudo-label quality.
• Empirical improvement over direct transfer, self-training baselines, and several state-of-the-art re-ID methods on person re-ID benchmarks (e.g., cross-dataset settings).
• Successful extension and validation of the framework to vehicle re-ID datasets, showing generality across domains.