Alen Adamyan

Smaller3D: Smaller Models for 3D Semantic Segmentation Using Minkowski Engine and Knowledge Distillation Methods

Alen Adamyan completed his fourth year at AUA Data Science in 2023 and simultaneously worked at Arloopa as an ML Engineer/Researcher for over a year, starting in 2022. During his work, he specialized in 3D (3D amateur auto rigging, for example), pushing the boundaries of this technology. Alen also focused on AR/VR projects, immersing himself in the exciting world of AR and VR. Additionally, he had the opportunity to work on projects involving other modalities, including AVSR and NLP. 


Description of the Talk:

There are various optimization techniques in the realm of 3D, including point cloud-based approaches that use mesh, texture, and voxels, which optimize how you store and calculate in 3D. These techniques employ methods such as feed-forward networks, 3D convolutions, graph neural networks, transformers, and sparse tensors. However, the field of 3D is one of the most computationally expensive fields, and these methods have yet to achieve their full potential due to their large capacity, complexity, and computation limits. This paper proposes the application of knowledge distillation techniques, especially for sparse tensors in 3D deep learning, to reduce model sizes while maintaining performance. The paper analyzes and proposes different loss functions, including standard methods and combinations of various losses, to simulate the performance of state-of-the-art models of different Sparse Convolutional NNs. The experiments are done on the standard ScanNet V2 dataset, and the researchers achieved around 2.6% mIoU difference with a 4 times smaller model and around 8% with a 16 times smaller model on the latest state-of-the-art spacio-temporal convents-based models.