| Property | Value |
|---|---|
| Parameter Count | 689M |
| License | CC BY-NC-SA 4.0 |
| Author | Naver |
| Paper | arXiv:2406.09756 |
| Architecture | ViT-Large encoder with ViT-Base decoder |
What is MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric?
MASt3R is a sophisticated image-to-3D model that specializes in grounding image matching in three-dimensional space. Developed by Naver, it represents a significant advancement in geometric 3D vision, utilizing a Vision Transformer (ViT) architecture with a Large encoder and Base decoder configuration.
Implementation Details
The model employs an asymmetric architecture with CatMLP+DPT head design, supporting multiple training resolutions (512x384, 512x336, 512x288, 512x256, 512x160). It's implemented in PyTorch and uses F32 tensor types for precise computations.
- Asymmetric architecture combining ViT-Large encoder with ViT-Base decoder
- CatMLP+DPT head for enhanced feature processing
- Multiple resolution support for flexible input handling
- Efficient implementation with 689M parameters
Core Capabilities
- Advanced 3D geometric vision processing
- High-quality image matching in 3D space
- Flexible resolution handling for various input sizes
- Robust feature extraction and matching
Frequently Asked Questions
Q: What makes this model unique?
The model's unique asymmetric architecture and CatMLP+DPT head design enable superior 3D vision capabilities while maintaining computational efficiency. Its ability to handle multiple resolutions makes it versatile for various applications.
Q: What are the recommended use cases?
This model is ideal for applications requiring precise 3D geometric vision, including 3D reconstruction, image matching in 3D space, and computer vision tasks requiring accurate spatial understanding.