Abstract: The demand for immersive and lifelike visual experiences has accelerated the development of imaging and display technologies capable of reproducing the real world with high fidelity. Among various approaches, light field imaging has emerged as a critical paradigm owing to its ability to capture both the spatial and angular characteristics of light rays in a scene. By recording the direction and intensity of light at each point, light fields enable post-capture refocusing, aperture modification, viewpoint synthesis and depth reconstruction, which are features that are fundamental to realistic 3D visualization. These properties have made light fields an essential component in next-generation autostereoscopic and computational multi-view displays which aim to provide glasses-free 3D experiences with natural motion parallax, wider field-of-view and greater depth perception.

However, the comprehensive nature of light fields introduces a critical challenge: the enormous volume of information generated. A single light field frame containing multi-view images can surpass gigabytes in size, resulting in significant bottlenecks in storage, transmission and real-time rendering. The problem becomes even more complex for dynamic light fields or light field videos, where temporal variations further multiply data volume. The central focus of this research is therefore to design efficient representation and compression schemes that can preserve the intrinsic structure of light fields while substantially reducing data size, and ensuring compatibility with computational multi-view and autostereoscopic display pipelines.

To address the data challenge, we explore integrated frameworks that effectively exploit spatial, angular and temporal correlations inherent in light field data. The proposed approaches provide compact and scalable representations that maintain flexibility across varying bitrates and adapt to different display configurations. The first part of the thesis focuses on the efficient compression of static light fields. We begin by formulating a multiplicative layered representation based on Block Krylov subspace approximation, which captures redundancy across multiple stacked layers through low-rank modeling. The factorized layers are subsequently encoded using High Efficiency Video Coding (HEVC) to remove residual redundancies. This framework achieves significant reductions in bitrate and computation time while preserving reconstruction fidelity.

Building upon this, a hierarchical hybrid model is proposed that integrates the Krylov subspace-based layered representation with Fourier Disparity Layer (FDL) analysis. By performing joint operations in both spatial and frequency domains, the scheme efficiently removes additional non-linear redundancies across sub-aperture images in multiple scanning patterns. This hierarchical approach further enhances scalability, enabling multi-rate reconstruction within a single unified framework. The next contribution introduces a Tucker decomposition–based hybrid framework using randomized TensorSketch, where the light field is modeled as a high-dimensional tensor. The proposed system performs single-pass, memory-efficient low-rank approximation and combines it with FDL-based frequency domain analysis to yield representations suitable for real-time streaming of large-scale light fields.

The static light field coding models are further extended through a focal stack–driven compression framework that employs hybrid Tucker–TensorSketch decomposition with vector quantization. By operating on a small set of differently focused images rather than the complete light field, this method achieves notable reductions in acquisition and computational cost while maintaining reconstruction fidelity. Collectively, these techniques establish a foundation for scalable, low-rank and display-adaptive light field compression, by progressively balancing efficiency and quality for autostereoscopic visualization.

The second part of this thesis discusses dynamic light fields or light field videos, where spatial, angular and temporal dimensions must be jointly represented and compressed. A hybrid Tucker–TensorSketch Vector Quantization (HTTSVQ) scheme is proposed to model dynamic light fields acquired through optimized coded aperture patterns. The low-rank approximated images are subsequently encoded using HEVC, allowing incremental streaming without storing entire sequences. This approach effectively handles temporal redundancy and demonstrates significant bitrate savings compared to conventional video codecs.

Next, a novel data-driven approach based on Dynamic Mode Decomposition (DMD) is introduced to represent dynamic light fields. The proposed model synchronously integrates aperture coding on the angular plane with pixel-wise exposure coding on the spatial plane. This formulation enables the embedding of light field video frames within an exposure time into a single coded image from which spatio-angular-temporal correlations are capitalized. Subsequent HEVC encoding ensures compact bitstreams with high reconstruction fidelity and computational efficiency. This work represents one of the first attempts to conceptualize light field videos as dynamical systems.

Through the frameworks proposed across both static and dynamic domains, we establish a unified strategy for light field representation and compression that is flexible, scalable and display-adaptive. Each method systematically explores low-rank and hybrid decomposition models by combining mathematical matrix techniques with data-driven learning to achieve structure-preserving representations. The proposed schemes demonstrate substantial improvements in compression ratio, memory efficiency, computational cost and rate–distortion performance while maintaining quality of reconstructions. Moreover, our models can additionally complement or enhance existing light field coding systems that support multi-view architectures. The results further highlight the feasibility of implementing these schemes in real-time or resource-constrained environments, paving the way for practical deployment in autostereoscopic display systems.

Overall, our research contributes to bridging the gap between high-dimensional light field capture and its efficient visualization in real-world 3D platforms. By addressing challenges in compact representation, scalable coding and reconstruction, the proposed frameworks advance the state of the art in light field compression. They enable new directions in light field imaging that can be integrated with next-generation AR/VR environments, holographic setups and multi-layered display technologies. The outcomes of this work have broader implications for computational photography, immersive visualization and high-dimensional video processing where the ability to represent and transmit richly detailed scenes remains a fundamental requirement.

The methods developed through this research thus hold promise not only for improving current autostereoscopic display systems, but also for inspiring future advances in 3D visual communication. As light field technologies continue to evolve, the insights and frameworks presented are expected to contribute meaningfully to the development of efficient representation and optimized coding solutions that will support the next generation of immersive media experiences.

Event Details
Title: Efficient Representation and Compression of Light Fields for Glasses-Free 3D Display Applications (PhD Viva Voce)
Date: March 25, 2026 at 02:00 PM
Venue: Google Meet (https://meet.google.com/edj-yobq-vrg)
Speaker: Ms. Joshitha R (EE19D701)
Guide: Dr. Mansi Sharma
Type: PHD seminar

Updated: