PaperlayerView exact revision on arXiv

Source-linked AI summary

IDM: An Intermediate Domain Module for Domain Adaptive Person Re-ID

Yongxing Dai, Jun Liu, Yifan Sun, Zekun Tong, Chi Zhang, Ling-Yu Duan

arXiv:2108.02413v1cs.CV

TL;DR

UDA re-ID must transfer labeled source-domain knowledge to an unlabeled target domain despite potentially extreme distribution differences. IDM models intermediate domains by mixing hidden source and target representations, and the method outperforms state-of-the-art approaches across common UDA re-ID tasks.

  • Problem

    UDA re-ID transfers labeled source-domain knowledge to improve discriminability on an unlabeled target domain, where source and target distributions may be extreme and their label spaces may not overlap.

  • Method

    IDM generates intermediate-domain representations on-the-fly by mixing hidden source and target representations with two domain factors, while bridge and diversity losses guide their placement and variation.

  • Results

    The method outperforms state-of-the-art methods by a large margin on all common UDA re-ID tasks.

  • Takeaways & Limitations

    Intermediate domains are used as a bridge to smoothly transfer source knowledge toward the target domain and improve target-domain discriminability.

  • Takeaways & Limitations

    The approach assumes that appropriate intermediate domains lie on the shortest geodesic path between the source and target domains.

Abstract

from arXiv · show

Unsupervised domain adaptive person re-identification (UDA re-ID) aims at transferring the labeled source domain's knowledge to improve the model's discriminability on the unlabeled target domain. From a novel perspective, we argue that the bridging between the source and target domains can be utilized to tackle the UDA re-ID task, and we focus on explicitly modeling appropriate intermediate domains to characterize this bridging. Specifically, we propose an Intermediate Domain Module (IDM) to generate intermediate domains' representations on-the-fly by mixing the source and target domains' hidden representations using two domain factors. Based on the "shortest geodesic path" definition, i.e., the intermediate domains along the shortest geodesic path between the two extreme domains can play a better bridging role, we propose two properties that these intermediate domains should satisfy. To ensure these two properties to better characterize appropriate intermediate domains, we enforce the bridge losses on intermediate domains' prediction space and feature space, and enforce a diversity loss on the two domain factors. The bridge losses aim at guiding the distribution of appropriate intermediate domains to keep the right distance to the source and target domains. The diversity loss serves as a regularization to prevent the generated intermediate domains from being over-fitting to either of the source and target domains. Our proposed method outperforms the state-of-the-arts by a large margin in all the common UDA re-ID tasks, and the mAP gain is up to 7.7% on the challenging MSMT17 benchmark. Code is available at https://github.com/SikaStar/IDM.

1. Introduction

UDA re-ID transfers labeled source-domain knowledge to an unlabeled target domain, but extreme distribution and label-space differences make direct adaptation difficult. IDM addresses this gap by generating constrained intermediate domains that bridge the source and target during training.

  • UDA re-ID transfers labeled source knowledge to improve discriminability on an unlabeled target domain despite potentially extreme distributions and non-overlapping label spaces.
  • Existing methods do not explicitly exploit information about similarity and dissimilarity between the source and target domains as a bridge.
  • The method models appropriate intermediate domains as lying along a shortest path between the source and target, enabling gradual knowledge transfer across the domain shift.
  • IDM mixes source and target hidden representations using two domain factors to generate intermediate-domain representations on-the-fly at any hidden network stage.
  • Bridge losses constrain intermediate-domain prediction and feature relationships, while diversity loss prevents domain factors from over-fitting to either extreme domain.
  • IDM outperforms state-of-the-art methods by a large margin across common UDA re-ID tasks.

2. Related Work

Prior UDA methods use adversarial learning, discrepancy metrics, image translation, fine-tuning, or joint training, while traditional manifold methods construct intermediate domains. IDM differs by explicitly learning intermediate representations to bridge extreme domains under constrained mixing rather than using unconstrained augmentation.

  • Unsupervised Domain Adaptation: UDA research includes adversarial and discrepancy-minimization methods, while traditional approaches bridge source and target data through intermediate domains on manifolds.
  • Unsupervised Domain Adaptative Person Re-ID: UDA person re-ID methods are commonly organized into GAN transferring, fine-tuning, and joint-training schemes.
  • Proposed approach: IDM mixes source and target hidden representations to create intermediate representations, feeds all domains through subsequent network stages, and applies re-ID and bridge-related losses.
  • Unsupervised Domain Adaptative Person Re-ID: Unlike fine-tuning and GAN-based approaches, IDM explicitly models bridging between extreme domains to support gradual adaptation.
  • Mixup and Variants: Unlike Mixup variants that use random interpolation for augmentation, IDM controls domain bridging with a dedicated module and specific losses.

3. Proposed Method

The method jointly trains source, target, and generated intermediate domains, using an IDM to mix hidden representations and bridge the two extreme domains. It constrains these intermediate domains through geodesic-distance relationships and diversity of domain factors.

  • Training Objective: The module jointly trains on source and target data while enforcing bridge losses on intermediate predictions and features and a diversity loss on domain factors.The diversity loss prevents the module from over-fitting to either extreme domain.
  • Intermediate Domain Module: IDM generates intermediate-domain representations on-the-fly by mixing source and target hidden representations, and it can be inserted after a hidden backbone stage.The module takes both domains’ hidden representations, derives two domain factors, and uses them for mixing during joint training.
  • Shortest Geodesic Path: Appropriate intermediate domains are defined as lying on the shortest geodesic path, satisfying d(Ps, Pinter) + d(Pt, Pinter) = d(Ps, Pt).Intermediate domains off this path introduce extra domain shifts because their distances to the source and target exceed the direct source-target distance.
  • Distance Properties: The domain factors are constrained to model proportional distances between intermediate and extreme domains, with as + at = 1 and endpoints recovering the source or target domain.When as approaches 1, the intermediate domain becomes more relevant to the source; when λ equals 0 or 1, the corresponding endpoint is recovered.
  • Losses: The bridge loss balances distances to the source and target according to the domain factors, while diversity maximizes factor variation within a mini-batch.If as exceeds at, the bridge loss penalizes the source-side distance more strongly; otherwise it emphasizes the target-side distance.

4. Experiments

Experiments evaluate IDM through component ablations, insertion-stage studies, backbone comparisons, mixup comparisons, and UDA re-ID benchmarks. IDM improves baselines and outperforms state-of-the-art methods, including on synthetic-to-real tasks.

  • Ablation Study: mAP/R1 improve by 4.7%/3.5% over Baseline2 and 5.2%/3.9% over Baseline1 on Market→Duke with full IDM.The strong baseline includes XBM, which is used for fair comparison with memory-bank-based state-of-the-art methods.
  • Ablation Study: mAP/R1 decrease 3.2%/1.8% without the prediction-space bridge loss and 1.5%/1.6% without the feature-space bridge loss.Using both bridge losses encourages intermediate domains to lie on the geodesic path between source and target domains.
  • Ablation Study: mAP decreases 2.7% without the diversity loss on Market→Duke, indicating its importance for the full IDM.The diversity loss is intended to prevent intermediate domains from over-fitting to either source or target domains.
  • Ablation Study: IDM performance declines as it is inserted after deeper ResNet-50 stages, so the experiments use stage-0 by default.The study reports that all insertion stages outperform the strong baseline, while deeper layers have weaker transferable ability.
  • Backbone and Mixup Comparisons: IDM scales to alternative backbones, with especially clear gains on the challenging MSMT dataset when using IBN-ResNet-50.The module is also described as compatible with AdaBN and CBN normalization techniques.
  • Comparison with the State-of-the-arts: On synthetic→real tasks, IDM exceeds SpCL mAP by 7.5%, 1.3%, and 7.6% on Market-1501, DukeMTMC-reID, and MSMT17, respectively.The paper also reports significant superiority over state-of-the-art methods across real→real and synthetic→real UDA re-ID settings.

5. Conclusion

The paper concludes that IDM uses intermediate domains to bridge source and target domains in UDA re-ID. Bridge and diversity losses constrain these domains, and experiments support the method’s effectiveness.

  • IDM generates intermediate domain representations to transfer source knowledge toward the unlabeled target domain.
  • Bridge losses position intermediate domains along an appropriate path between source and target domains.
  • A diversity loss prevents intermediate domains from over-fitting to either extreme domain.

A. Overall Training and Testing Procedure

The training procedure uses IDM during optimization but discards it during testing. Testing relies on the backbone’s normalized features.

  • IDM is used only during training and discarded for testing.
  • The overall training procedure uses XBM as a memory bank in the Strong Baseline implementation.
  • The testing procedure uses L2-normalized backbone features.

B. Implementation Details

The implementation uses a ResNet-50 backbone with domain-specific batch normalization and standard image preprocessing. IDM is configured through fully connected layers and a reduction ratio.

  • ResNet-50 pretrained on ImageNet is used as the backbone with domain-specific batch normalization.
  • Images are resized to 256×128 and augmented with random flipping, cropping, and erasing.
  • IDM uses fully connected layers whose dimensions depend on the feature-channel count and reduction ratio r.

B.1. Clustering on the Target Domain

The supplementary implementation describes an IDM training pipeline with source-target batches, pseudo-label clustering, feature memory, and configurable module hyperparameters.

  • Target samples receive pseudo labels through DBSCAN clustering using Jaccard distance and k-reciprocal neighborhoods.
  • The training algorithm initializes ImageNet-pretrained ResNet-50, an empty XBM queue, and IDM after a selected backbone stage.
  • The pipeline jointly produces source, target, and intermediate-domain features and predictions, optionally mines hard negatives from XBM, and optimizes the overall loss.

B.3. Implementation of XBM

The Strong Baseline uses a memory bank implemented as a queue to mine hard negatives from both source and target domains for triplet loss.

  • The memory bank is maintained and updated as a queue during training.For each mini-batch, features are enqueued to support hard-negative mining.

C. Additional Experimental Results

On two additional real →real UDA re-ID tasks, the method outperforms state-of-the-art methods by a large margin and significantly outperforms them across existing benchmarks.

  • The method significantly outperforms state-of-the-art methods across all existing UDA re-ID benchmarks.The additional tasks use MSMT17 as source and Market-1501 or DukeMTMC-reID as targets.
  • The evaluated additional tasks are MSMT17 → Market-1501 and MSMT17 → DukeMTMC-reID.

D. Parameter Analysis

The parameter analysis examines the IDM reduction ratio and memory-bank size, while the reported loss-weight figures assess additional hyperparameters.

  • Figures 6–8 evaluate performance across the loss weights µ1, µ2, and µ3.
  • A larger IDM reduction ratio slightly lowers performance because it makes the IDM harder to learn, so r is set to 2.The reduction-ratio study plugs IDM after stage-0 of ResNet-50.
  • The memory ratio RM controls memory-bank size, with RM = 0 corresponding to the Naive Baseline and RM = 1 to the Strong Baseline.

D.3. The loss weight µ1

The loss-weight analysis tunes the prediction-space bridge loss, feature-space bridge loss, and diversity loss to balance intermediate-domain supervision and source-target learning.

  • Performance improves as µ1 increases to 0.7, then slightly degrades from 0.7 to 0.9, so µ1 is set to 0.7.µ1 weights the bridge loss on intermediate domains’ prediction space.
  • A feature-space bridge-loss weight near µ2 = 0.1 improves performance, whereas a large µ2 slightly degrades it by emphasizing intermediate domains over source and target domains.
  • The diversity-loss weight is set to µ3 = 1.0 for the remaining UDA re-ID experiments.
Loading 2108.02413v1…