Minifying photometric stereo via knowledge distillation-based feature translation.

Area of Science:

• Computer Vision
• Computer Graphics
• Computational Imaging

Background:

• Photometric stereo (PS) is crucial for estimating surface normals but typically requires numerous images, leading to significant capture effort and computational expense.
• Existing methods face challenges in accuracy and efficiency when dealing with a limited number of input images.

Purpose of the Study:

• To develop a robust photometric stereo method that achieves high-quality surface normal estimation with a minimal number of input images.
• To reduce the computational cost and image acquisition requirements of traditional photometric stereo techniques.

Main Methods:

• Proposed a novel feature translation module (FTM) integrated into a baseline PS network to enrich sparse feature information.
• Employed distillation loss, using a teacher PS network trained on extensive data, to supervise FTM activations.
• Demonstrated applicability to both calibrated and uncalibrated photometric stereo scenarios.

Main Results:

• The proposed PS network successfully estimates surface normals with high accuracy using very few input images.
• Achieved output quality comparable to a teacher PS network trained on a large dataset.
• Validated effectiveness across various input conditions, including limited image availability.

Conclusions:

• The developed method offers a significant advancement in photometric stereo, enabling accurate surface normal estimation with reduced data and computational load.
• This approach enhances the practicality and efficiency of PS applications in diverse scenarios.
• The feature translation module effectively addresses the challenge of scarce information in low-image-count PS.