Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Less is More: Learning from Synthetic Data with Fine-grained Attributes for Person Re-Identification (2109.10498v3)

Published 22 Sep 2021 in cs.CV

Abstract: Person re-identification (re-ID) plays an important role in applications such as public security and video surveillance. Recently, learning from synthetic data, which benefits from the popularity of synthetic data engine, has attracted great attention from the public eyes. However, existing datasets are limited in quantity, diversity and realisticity, and cannot be efficiently used for re-ID problem. To address this challenge, we manually construct a large-scale person dataset named FineGPR with fine-grained attribute annotations. Moreover, aiming to fully exploit the potential of FineGPR and promote the efficient training from millions of synthetic data, we propose an attribute analysis pipeline called AOST, which dynamically learns attribute distribution in real domain, then eliminates the gap between synthetic and real-world data and thus is freely deployed to new scenarios. Experiments conducted on benchmarks demonstrate that FineGPR with AOST outperforms (or is on par with) existing real and synthetic datasets, which suggests its feasibility for re-ID task and proves the proverbial less-is-more principle. Our synthetic FineGPR dataset is publicly available at https://github.com/JeremyXSC/FineGPR.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (7)
  1. Suncheng Xiang (27 papers)
  2. Guanjie You (3 papers)
  3. Mengyuan Guan (4 papers)
  4. Hao Chen (1006 papers)
  5. Binjie Yan (2 papers)
  6. Ting Liu (329 papers)
  7. Yuzhuo Fu (24 papers)
Citations (25)

Summary

Learning from Synthetic Data with Fine-grained Attributes for Person Re-Identification

Person re-identification (re-ID) is a pivotal task in computer vision, specifically within applications like public security and video surveillance, where identifying individuals across different camera views is imperative. The paper "Less is More: Learning from Synthetic Data with Fine-grained Attributes for Person Re-Identification" introduces an innovative approach to address inherent challenges in the re-ID domain concerning data limitations, diversity, and privacy concerns.

Introduction to FineGPR Dataset

The authors present the Fine-grained GTA Person Re-ID (FineGPR) dataset, which leverages the graphics capabilities of the GTA5 game engine to synthesize a vast dataset featuring 2,028,600 images of 1,150 identities. Unlike existing datasets, FineGPR is meticulously annotated with fine-grained attributes at both the environment and individual levels, enabling rich data diversity. This includes attributes like viewpoint, weather conditions, illumination variations, and ID-level specifications, providing a robust framework for training re-ID models without the ethical and privacy issues associated with real-world data.

Methodology: AOST Pipeline

To maximize the utility of the FineGPR dataset, the paper introduces the Attribute Optimization and Style Transfer (AOST) pipeline. AOST dynamically adjusts the attribute distributions from the synthetic domain to closely mirror those found within real-world environments, enhancing the dataset's applicability. The pipeline consists of two stages:

  1. Attribute Optimization: Using a tree boosting system (XGBoost), combined with Particle Swarm Optimization (PSO), the pipeline optimizes the selection of samples within FineGPR to align with the target attribute distribution of a real dataset.
  2. Style Transfer: The AOST employs Generative Adversarial Networks (GANs) to resolve the domain gap between synthetic and real-world images, utilizing the newly established high-resolution MSCO dataset to produce images with photorealistic attributes.

Numerical Results and Implications

The paper provides a thorough evaluation of the FineGPR dataset and the AOST methodology against state-of-the-art benchmarks. Results demonstrate that FineGPR, coupled with AOST, achieves competitive performance in re-ID tasks across various standard datasets (Market-1501, MSMT17, and CUHK03), sometimes surpassing results obtained with real datasets. For example, FineGPR with AOST notably improved Rank-1 accuracy on Market-1501 and CUHK03 compared to direct training on RandPerson.

Practical Implications and Future Directions

The development of the FineGPR dataset alongside the AOST strategy addresses key limitations of real-world data collection constraints, offering an alternative path for model training without sacrificing model performance or ethical concerns. This work exemplifies how synthetic data, when intelligently utilized and optimized, can considerably contribute to tasks traditionally reliant on real data.

Future research can explore expanding the synthesis engine of FineGPR to include more complex scenarios and attributes, further refining the AOST pipeline with more sophisticated optimization techniques, and investigating the application of this approach within other areas of computer vision, such as crowd behavior analysis and event detection.

In conclusion, this paper showcases a forward-thinking approach to person re-identification by leveraging synthetic data, potentially reshaping datasets' role in privacy-sensitive machine learning scenarios.

Github Logo Streamline Icon: https://streamlinehq.com

GitHub