Reproduce “MobileViTv3: Mobile-Friendly Vision Transformer with Simple and Effective Fusion of Local, Global and Input Features” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shakti N. Wadekar, Abhishek Chaurasia
Researchers from Purdue University and Micron Technology Inc. enhanced MobileViT architectures by developing MobileViTv3, an improved block that simplifies and optimizes the fusion of local, global, and input features. This led to consistent accuracy gains across classification, segmentation, and detection tasks, such as a 2.0% increase in ImageNet-1K Top-1 accuracy for MobileViTv3-XXS over MobileViTv1-XXS, while maintaining competitive computational costs.
Key insights
- Replacing a 3x3 convolution with a 1x1 convolution in the fusion block simplifies the learning task and significantly improves accuracy and scaling efficiency.
- Fusing features from the local representation block directly with global features, rather than input features, proves more effective for integrating complementary information.
- Incorporating the original input features as a residual connection after the fusion block further enhances model optimization and overall accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.