Version 1
: Received: 28 March 2024 / Approved: 28 March 2024 / Online: 28 March 2024 (12:54:44 CET)
How to cite:
Miya, R.; Kawaguchi, T.; Saito, T. Balancing Speed and Precision: Lightweight and Accurate Depth Estimation for Light Field Image. Preprints2024, 2024031771. https://doi.org/10.20944/preprints202403.1771.v1
Miya, R.; Kawaguchi, T.; Saito, T. Balancing Speed and Precision: Lightweight and Accurate Depth Estimation for Light Field Image. Preprints 2024, 2024031771. https://doi.org/10.20944/preprints202403.1771.v1
Miya, R.; Kawaguchi, T.; Saito, T. Balancing Speed and Precision: Lightweight and Accurate Depth Estimation for Light Field Image. Preprints2024, 2024031771. https://doi.org/10.20944/preprints202403.1771.v1
APA Style
Miya, R., Kawaguchi, T., & Saito, T. (2024). Balancing Speed and Precision: Lightweight and Accurate Depth Estimation for Light Field Image. Preprints. https://doi.org/10.20944/preprints202403.1771.v1
Chicago/Turabian Style
Miya, R., Tatsuya Kawaguchi and Takushi Saito. 2024 "Balancing Speed and Precision: Lightweight and Accurate Depth Estimation for Light Field Image" Preprints. https://doi.org/10.20944/preprints202403.1771.v1
Abstract
With the progression of AI, embedding advanced AI technologies into small robotics and mobile devices has become essential, driving research towards lightweighting AI models. Our study enhances the EPINET depth estimation model for light field images, aiming for compactness and faster inference while preserving accuracy. We conducted two-step experiments aimed at enhancing inference efficiency: Initially, by adjusting input streams and convolution layers, we simplified the CNN model, achieving faster inference times at the cost of reduced accuracy. To address this reduction in accuracy, we then applied knowledge distillation, allowing the simplified model to learn from the original model’s more complex patterns. In our quantitative experiments using two error metrics, MSE (Mean Squared Error) and BadPix, we identified optimal knowledge positions and evaluated the required complexity for the student model. As a result, our method improved MSE by 21% and BadPix by 14% compared to training without it. Furthermore, the student model achieved an inference speed 13% faster than the teacher model and surpassed its accuracy by 10% in MSE. Additionally, we demonstrated that repeatedly applying our approach could further enhance both model compactness and accuracy.
Computer Science and Mathematics, Artificial Intelligence and Machine Learning
Copyright:
This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
