Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
41 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
41 tokens/sec
o3 Pro
7 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Adaptive Exploration for Unsupervised Person Re-Identification (1907.04194v2)

Published 9 Jul 2019 in cs.CV

Abstract: Due to domain bias, directly deploying a deep person re-identification (re-ID) model trained on one dataset often achieves considerably poor accuracy on another dataset. In this paper, we propose an Adaptive Exploration (AE) method to address the domain-shift problem for re-ID in an unsupervised manner. Specifically, in the target domain, the re-ID model is inducted to 1) maximize distances between all person images and 2) minimize distances between similar person images. In the first case, by treating each person image as an individual class, a non-parametric classifier with a feature memory is exploited to encourage person images to move far away from each other. In the second case, according to a similarity threshold, our method adaptively selects neighborhoods for each person image in the feature space. By treating these similar person images as the same class, the non-parametric classifier forces them to stay closer. However, a problem of the adaptive selection is that, when an image has too many neighborhoods, it is more likely to attract other images as its neighborhoods. As a result, a minority of images may select a large number of neighborhoods while a majority of images have only a few neighborhoods. To address this issue, we additionally integrate a balance strategy into the adaptive selection. We evaluate our methods with two protocols. The first one is called "target-only re-ID", in which only the unlabeled target data is used for training. The second one is called "domain adaptive re-ID", in which both the source data and the target data are used during training. Experimental results on large-scale re-ID datasets demonstrate the effectiveness of our method. Our code has been released at https://github.com/dyh127/Adaptive-Exploration-for-Unsupervised-Person-Re-Identification.

Adaptive Exploration for Unsupervised Person Re-identification

The paper explores the challenges and solutions associated with unsupervised person re-identification (re-ID), particularly addressing the domain-shift problem evident when models trained on one dataset underperform on another. The authors propose an innovative Adaptive Exploration (AE) method to tackle this issue.

The fundamental challenge in person re-ID is the domain bias stemming from differences in cloth styles, camera viewpoints, and other environmental variables. Traditional methods rely heavily on labeled data to fine-tune models for different datasets, but this can be resource-intensive and impractical for large-scale or new datasets. Instead, this paper focuses on leveraging unlabeled data to train models in an unsupervised manner, addressing the domain-shift without extensive labeling efforts.

Methodology

The AE method operates by maximizing inter-image distances while minimizing distances between similar images. This approach involves treating each person image as an individual class and employs a non-parametric classifier with feature memory. The feature memory retains features of all person images in the target domain, updating iteratively as the model trains. The process encourages distinct separation between different classes (individual images) in the feature space and enforces close clustering of similar images. An adaptive strategy selects neighborhoods for each image, defined by a similarity threshold, thus ensuring reliable clustering.

A noticeable challenge arises when an image attracts an excessive number of neighborhoods, leading to significant imbalances. In this situation, fewer images select neighborhoods predominantly, skewing the dataset. To counteract this, the AE approach integrates a balance strategy which adjusts the weight of the loss function, maintaining a fair distribution.

Experimental Evaluation

The authors conduct their evaluation on multiple large-scale person re-ID datasets using two protocols: a "target-only re-ID," employing only the unlabeled target data for model training, and a "domain adaptive re-ID," utilizing both source and target data. Their findings confirm the AE method's effectiveness, showcasing superior performance compared to state-of-the-art methods.

Numerical Outcomes

The experimental results exhibit notable improvements, with the AE method achieving a significant increase in both rank-1 accuracy and mean average precision (mAP) across different datasets. For instance, when tested on Market-1501 and DukeMTMC-reID, the AE method outperforms existing methods, demonstrating the successful handling of domain shifts without labeled data.

Implications and Future Directions

The AE method introduces a potentially transformative approach in reducing the dependence on plentiful labeled data in person re-ID. By capitalizing on domain adaptation and unsupervised learning, this research broadens the operational scope of re-ID systems with promising implications for security and surveillance operations, where real-time and adaptive model performance is crucial.

Potential avenues for future research include enhancing adaptive selection mechanisms further, reducing computational overhead, and integrating this methodology with other unsupervised learning techniques to improve robustness and accuracy further. Exploring these methods under various environmental conditions and novel dataset structures can offer richer insights and facilitate more comprehensive model applications.

In conclusion, the proposed Adaptive Exploration method provides a critical step toward more adaptable, efficient, and scalable approaches for unsupervised person re-identification, and sets a solid foundation for future innovations in domain adaptation and unsupervised learning.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Yuhang Ding (5 papers)
  2. Mingliang Xu (86 papers)
  3. Yi Yang (855 papers)
  4. HeHe Fan (46 papers)
Citations (128)