Model compression and inference optimization for drone research in SOTI Aerospace Research Team, supervised by Prof. Nandita Vijaykumar and Dr. Sunil Jacob.
| Previous Model (SegFormer) |
Current Model (EfficientViT with Model Compression) |
Δ | |
|---|---|---|---|
| mloU | 46.51 | 50.48 | + 8.5% |
| Latency (s) | 0.0927 | 0.0310 | - 66.56% |
| FPS | 10.79 | 32.26 | + 198.98% |
- Designed a model compression pipeline to accelerate vision models, especially for real-time indoor scene segmentation on drone’s companion computers
- Introduced a novel feature-based knowledge distillation method for image classification and semantic segmentation, combining self-supervised learning with reused classifier
- Attained a remarkable 79.91% classification accuracy on CIFAR-100 with ResNet-8x4 student model, surpassing state-of-the-art approaches by 1.83% (e.g., SimKD, DIST, and DKD)
- Deployed SegFormer on the Jetson Xavier board using OpenMMLab and PyTorch, achieving a 16.14% boost in mIoU and a 43.8% reduction in model size compared to the previous segmentation model
- Developed a distillation codebase for MMSegmentation models, simplifying the implementation of distillation loss functions and training pipelines, thus facilitating knowledge distillation for semantic segmentation research
- Optimized semantic segmentation models for drone deployment on the Jetson Orin using inference optimization and model quantization, leading to a notable 66.56% decrease in inference latency
Yuxuan (Reacher) Zhang
Master of Science in Applied Computing (MScAC)
My research interests include diffusion models, 3D content generation, knowledge distillation and model compression.