Introduction
Compressive Imaging (CI) is a powerful technique for capturing high-dimensional data—such as 3D video or 4D medical scans—using low-dimensional sensors. However, reconstructing the original signal from these compressed measurements remains a significant challenge. Traditional approaches often struggle with either the need for massive supervised datasets, the "spectral bias" of implicit neural representations (INRs) leading to blurry results, or the exponential scaling of grid-based methods like InstantNGP.
The authors introduce GridTD (Tensor Decomposed multi-resolution Grid encoding), an unsupervised framework designed to overcome the "curse of dimensionality." GridTD combines the high-frequency detail preservation of multi-resolution grids with the efficiency of tensor decomposition, enabling high-quality reconstruction with linear scaling in parameters and computation.
Methodology: The GridTD Framework
The core innovation of GridTD lies in its unique architecture that avoids the exponential memory growth typical of high-dimensional grids.
The framework consists of several key components:
- Coordinate Decomposition: Input coordinates are decomposed into $D$ independent 1D coordinates.
- 1D Multi-resolution Grids: Instead of a single $D$-dimensional grid, GridTD employs lightweight 1D grids for each dimension. Features are interpolated along these lines, significantly reducing memory overhead.
- Tensor Fusion: Inspired by Canonical Polyadic (CP) decomposition, these 1D features are fused via element-wise products to form a unified feature vector.
- Temporal Affine Adapter: For video tasks, a specialized module learns motion patterns (rotation, scaling, translation) to effectively reuse temporal features across frames.
- PnP-ADMM Integration: The model is optimized within a Plug-and-Play Alternating Direction Method of Multipliers (PnP-ADMM) framework, ensuring spatial-temporal smoothness and principled reconstruction.
Experimental Evaluation
GridTD was evaluated across three challenging high-dimensional imaging tasks:
- Video Snapshot Compressive Imaging (SCI): Reconstructing grayscale benchmark videos (e.g., Aerial, Vehicle, Kobe) at $256 \times 256 \times 8$ resolution.
- Spectral SCI: Recovering 31 spectral bands from a single 2D measurement using the KAIST hyperspectral dataset.
- Dynamic MRI: Cardiac cine imaging using the OCMR dataset to simulate various scan acceleration factors.
Key Results and Insights
GridTD demonstrates a significant advantage over existing methods, particularly as the dimensionality of the data increases.
Efficiency and Scalability
The most critical finding is the shift from exponential to linear scaling. While InstantNGP's parameter count grows as $O(N^D)$, GridTD scales as $O(D)$. In a 3D toy task, GridTD achieved superior results using only 0.02M parameters—a fraction of InstantNGP’s 3.12M—while running nearly 30x faster (1.31s vs 38.05s).
Reconstruction Quality
- Video SCI: GridTD improved average PSNR by approximately 1 dB over previous unsupervised state-of-the-art methods.
- Spectral SCI: With an average PSNR of 36.99 dB, it outperformed the best unsupervised competitor (LRSDN) and even surpassed some supervised transformer-based models.
- MRI: In cardiac imaging, GridTD was over 50% faster than InstantNGP and provided better preservation of structural edges during heart motion.
Theoretical Stability
The authors proved that GridTD’s Lipschitz constant scales linearly with dimension, whereas InstantNGP’s scales exponentially. This theoretical foundation explains GridTD’s increased stability and superior generalization in high-dimensional spaces.
Conclusion and Future Work
GridTD represents a major advancement in unsupervised compressive imaging. By leveraging tensor decomposition, it effectively bypasses the curse of dimensionality, offering a fast, memory-efficient, and high-fidelity solution for complex reconstruction tasks.
While the per-image optimization process remains a limitation compared to the instant inference of supervised models, the unsupervised nature of GridTD makes it invaluable for domains where ground-truth data is scarce. Future research will explore extending this framework to even higher-dimensional tasks, such as 4D+ super-resolution and hyperspectral-multispectral fusion.
Summary updated