Effectively lossless subspace appearance model compression using prior information

Yang Li, Xiaoming Tao, Jianhua LuTsinghua 2015 · DOI

Abstract

Subspace appearance models are widely used in computer vision and image processing tasks to compactly represent the appearance variations of target objects. In order to ensure algorithm performance, they are typically stored in high-precision formats; this results in a large storage footprint, rendering redistribution costly and difficult. Since for most image and vision applications, pixel values are quantized to 8 bits by the acquisition apparatuses, we show that it is possible to construct a fixed-width, effectively lossless representation of the bases vectors, in the sense that reconstructions from the original bases and from the quantized bases never deviate by more than half of the quantization step-size. In addition to directly applying this result to losslessly compress individual models, we also propose an algorithm to compress appearance models by utilizing prior information on the modeled objects in the form of prior appearance subspaces. Experiments conducted on the compression of person-specific face appearance models demonstrate the effectiveness of the proposed algorithms.

Introduction

Subspace appearance models, typically constructed via Principal Component Analysis (PCA), are fundamental tools in computer vision for representing objects with significant appearance variations, such as human faces. To maintain high reconstruction accuracy and numerical stability, these models are traditionally stored in high-precision formats (e.g., 64-bit floating-point). However, this leads to a substantial storage footprint, complicating the distribution of models to edge devices like smartphones. This paper addresses the discrepancy between the high precision of the models and the standard 8-bit quantization of the images they process, proposing an "effectively lossless" compression framework.

Methodology

Effectively Lossless Quantization (ELQ)

The authors define a representation as "effectively lossless" if the reconstruction from the compressed basis vectors deviates from the original high-precision model by no more than half of a pixel's quantization step (0.5 intensity levels). By formulating an optimization problem using Lagrange Multipliers, the authors derive an optimal quantization step for each basis vector. This ensures that bits are allocated based on the "importance" of each vector, determined by the standard deviation of its coefficients.

General Class Model (GCM) as Prior Knowledge

For Object-Specific Models (OSMs), such as a model of a specific individual's face, the authors propose leveraging prior knowledge. A "General Class Model" (GCM) is trained on a broad dataset (e.g., the Multi-PIE face database). The specific OSM is then projected onto this GCM, and only the projection coefficients and the residual error are stored. Since the OSM is highly correlated with the general class, the residuals exhibit low entropy and are highly compressible.

Key Insights: The Optimal Model Size

A critical finding is the trade-off between the complexity of the prior model and the resulting compression gains. While increasing the number of basis vectors in the GCM reduces the entropy of the residuals, it also increases the storage overhead of the GCM itself.

Figure 4: Total compressed size of 313 face models as a function of the number of GCM basis vectors used. The curve demonstrates a clear minimum, indicating an optimal balance between prior knowledge overhead and residual compression.

As shown in Figure 4, there is a mathematically optimal number of GCM basis vectors that minimizes the total storage size for a given set of objects. This "sweet spot" prevents the overhead of a massive GCM from outweighing the savings gained in residual entropy.

Experimental Results

The proposed methods were evaluated against standard compression tools and on various datasets, including the Multi-PIE database and common video sequences like "Akiyo" and "Miss America."

Performance on Standard Sequences

The ELQ method significantly outperforms general-purpose compression algorithms. For the "Akiyo" sequence, ELQ achieved a compressed size of 1.44 MB, compared to 5.31 MB for LZMA.

Figure 5: Compressed size (MB) for standard video sequences (Akiyo, Franck, Miss America) using varying percentages of GCM basis vectors. The 0% bars represent independent ELQ compression, which already provides substantial gains over standard tools.

Gains with Prior Knowledge

When applying the GCM-based approach to face models, the authors observed up to a 5-fold increase in the compression ratio compared to state-of-the-art methods. This efficiency stems from the sparsity of the residuals after GCM prediction; because the specific face is very similar to the general face model, the residuals have very low entropy and can be compressed much more efficiently than raw data.

Conclusion

This work demonstrates that by considering the 8-bit nature of digital imaging and utilizing prior class information, subspace appearance models can be compressed drastically without any perceptible loss in quality. Future work will explore formal rate-distortion trade-offs for scenarios where lossy compression is acceptable.

Summary updated

Abstract

Subspace appearance models are widely used in computer vision and image processing tasks to compactly represent the appearance variations of target objects. In order to ensure algorithm performance, they are typically stored in high-precision formats; this results in a large storage footprint, rendering redistribution costly and difficult. Since for most image and vision applications, pixel values are quantized to 8 bits by the acquisition apparatuses, we show that it is possible to construct a fixed-width, effectively lossless representation of the bases vectors, in the sense that reconstructions from the original bases and from the quantized bases never deviate by more than half of the quantization step-size. In addition to directly applying this result to losslessly compress individual models, we also propose an algorithm to compress appearance models by utilizing prior information on the modeled objects in the form of prior appearance subspaces. Experiments conducted on the compression of person-specific face appearance models demonstrate the effectiveness of the proposed algorithms.
BibTeX
@inproceedings{li2015effectively,
  author = {Yang Li and Xiaoming Tao and Jianhua Lu},
  title = {{Effectively lossless subspace appearance model compression using prior information}},
  booktitle = {Tsinghua},
  year = {2015},
  doi = {10.1109/tst.2015.7173455},
}

Loading related content...

Figure