Jongmin presented a paper at NeurIPS 2024 in Vancouver

Congratulations to Jongmin Lee for presenting his paper on 3D Pose Estimation at NeurIPS 2024, held in Vancouver, Canada.

The paper addresses one of the central challenges in 3D vision: accurately determining object orientations from a single image. Jongmin’s work introduces a novel approach that leverages SO(3)-equivariant pose harmonics in the frequency domain, enabling consistent and robust pose estimation under arbitrary rotations. This method avoids the pitfalls of traditional spatial parameterizations, such as discontinuities in Euler angles and singularities in quaternions.

Through extensive evaluation, the model demonstrates state-of-the-art performance on key benchmarks like ModelNet10-SO(3) and PASCAL3D+, showing significant improvements in accuracy, robustness, and data efficiency.

Research Highlights

• Introduced a frequency-domain Wigner-D coefficient regression framework compatible with spherical CNNs.
• Achieved strong generalization across multiple pose estimation datasets.
• Demonstrated efficiency in training with reduced data requirements.

About the Event

NeurIPS 2024 (Conference on Neural Information Processing Systems) is one of the world’s premier venues for machine learning and AI research, bringing together researchers and practitioners to share cutting-edge ideas. Jongmin’s presentation marks an important contribution from the computer vision community.