Distributed Manifold Hashing for Image Set Classification and Retrieval

Xiaobo Shen, Peizhuo Song, Yun-Hao Yuan, Yuhui ZhengAAAI 2024 · DOI

TL;DR

This paper introduces Distributed Manifold Hashing (DMH), a novel method for efficiently classifying and retrieving image sets that are distributed across multiple locations. By modeling these distributed sets as a graph and learning compact hash codes, DMH significantly reduces computational and storage costs while maintaining competitive accuracy, addressing a critical gap in current image set analysis.

What Problem Does This Paper Address?

Traditional image set classification and retrieval methods typically assume that all images belonging to a set are stored in a single location. This assumption breaks down in real-world scenarios, such as surveillance systems, where image data is often distributed across various devices or servers. Gathering all this data into one place for processing is computationally expensive and impractical for large-scale applications. Furthermore, existing methods often require significant storage for image set representations, posing another challenge for efficiency. This paper tackles the problem of distributed image set analysis and the need for efficient, compact representations in such settings.

Methods

The core of DMH lies in its ability to handle distributed data and learn compact representations. Here's a breakdown of the key techniques:

  • Riemannian Manifold Representation: Image sets are first represented using Riemannian manifolds, specifically SPD (Symmetric Positive Definite) or Grassmann manifolds. These manifolds are well-suited for capturing the intrinsic geometric structure of image sets.
  • Kernel Learning: To further capture nonlinear relationships, a kernel trick is applied to map these manifold representations into a high-dimensional Hilbert space.
  • Distributed Hashing: The crucial innovation is learning compact hash codes for each image set. This is formulated as a distributed learning problem.
  • Local Consistency Constraint: DMH imposes a local consistency constraint on global variables among neighboring nodes in the network. This allows for parallel optimization across multiple nodes, significantly speeding up training.
  • Augmented Lagrangian and ADMM: The optimization problem is non-convex and is solved using the Alternating Direction Method of Multipliers (ADMM) with an augmented Lagrangian formulation. This approach effectively decomposes the problem into subproblems that can be optimized locally.
  • Discrete Cyclic Coordinate Descent (DCC): For optimizing the binary hash codes, a DCC strategy is employed, which iteratively updates each bit of the hash code to find a closed-form solution.

What's new here is the combination of Riemannian manifold learning with hashing in a distributed setting, specifically designed to address the challenges of decentralized data and efficiency.

Datasets

The paper evaluates DMH on three benchmark image set datasets:

  • BBT: Contains 4,667 image sets with 15 classes.
  • PB: Contains 9,435 image sets with 20 classes.
  • YTC: Contains 1,856 image sets with 47 classes.

These datasets are standard for image set analysis and are used to compare DMH against existing state-of-the-art methods. The paper does not introduce new benchmarks.

Key Insights and Ideas

  • Distributed Learning is Crucial for Real-World Scenarios: The paper highlights that assuming centralized data is a significant limitation for practical applications. DMH's distributed approach directly addresses this.
  • Hashing for Efficiency: The core idea of using hashing to create compact representations is a powerful way to tackle storage and computational bottlenecks in large-scale image set tasks.
  • Graph-Based Modeling for Distributed Data: Representing distributed image sets as a connected graph allows for a structured approach to learning and optimization across nodes.
  • Parallelization Significantly Boosts Training Speed: The ability to optimize locally and communicate constraints across neighbors leads to substantial reductions in training time, as demonstrated by the experiments with varying numbers of nodes.
  • Binary Hashing Can Outperform Continuous Representations: Surprisingly, the paper shows that even with the binary constraint of hashing, DMH can achieve competitive or superior performance compared to its continuous variant, suggesting that the learned discriminative features are more important than the continuous nature of the representation itself.

Key Results

  • Accuracy: DMH-SPD (using SPD manifold) achieves state-of-the-art classification accuracies, outperforming the best baselines by up to 5.31% on the PB dataset. DMH-GM (using Grassmann manifold) also shows strong performance.
  • Retrieval: DMH-SPD and DMH-GM achieve the highest mean Average Precision (mAP) scores for image set retrieval across the datasets.
  • Efficiency (Storage): DMH's hash code representation requires significantly less storage (e.g., 64 times less than continuous features) compared to original features, manifold features, or kernel matrices.
  • Efficiency (Time): Hashing methods, including DMH, are substantially faster for distance calculation than traditional methods. DMH is nearly 8 times faster than GDA/CDL and up to 700 times faster than PML on the BBT dataset.
  • Scalability: Training time decreases significantly as the number of nodes increases, with parallel training across 8 nodes being approximately 10 times faster than centralized training on a single node.

Limitations and Future Work

  • Parameter Sensitivity: The performance of DMH is sensitive to the trade-off parameters $\alpha$ and $\nu$. The paper suggests that $\alpha$ should be less than 0.1 and $\nu$ larger than 100 for good performance.
  • Future Work: The paper's conclusion suggests that DMH is a promising direction for handling distributed image sets efficiently. Future work could explore more advanced distributed learning techniques or investigate different manifold representations. The authors also mention that their continuous variant (DMH-C) still performs well, suggesting potential for hybrid approaches.

Summary updated

BibTeX
@inproceedings{shen2024distributed,
  author = {Xiaobo Shen and Peizhuo Song and Yun-Hao Yuan and Yuhui Zheng},
  title = {{Distributed Manifold Hashing for Image Set Classification and Retrieval}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i5.28282},
}
Figure