DepthAnything/Depth-Anything-V2

Features

• Monocular Depth Estimators - Provides a foundation model that infers three-dimensional spatial depth from single two-dimensional image inputs.
• Computer Vision Models - Provides a large-scale pre-trained neural network designed for general purpose spatial understanding and depth perception.
• Depth Estimation - Calculates absolute distance measurements in indoor and outdoor scenes using scale-aware models.
• Metric - Calculates absolute distance measurements for indoor and outdoor scenes using specialized scale-aware models.
• Relative - Produces a relative depth map from a single input image using pre-trained foundation models.
• Temporal Video - Generates depth maps for video sequences while maintaining temporal consistency across frames.
• Video Depth Frameworks - Provides a framework for generating temporally consistent depth maps across sequential video frames.
• Relative-to-Metric Depth Scaling - Translates dimensionless relative depth maps into absolute distance measurements using scale-aware model variants.
• Spatial Understanding - Extracts fine-grained geometric information from images to perceive the layout of a physical space.
• Metric Depth Mapping - Determines absolute physical distance between the camera and objects in indoor or outdoor environments.
• Video Depth Analysis - Generates consistent depth maps across video frames to understand the three dimensional structure of moving scenes.
• Temporal Prediction Smoothing - Ensures depth predictions remain stable and smooth across consecutive video frames to reduce jitter.
• Transformer Encoders - Uses a vision transformer architecture to extract global context and high-resolution spatial features.
• Unsupervised Pre-training - Implements unsupervised pre-training on massive unlabeled datasets to learn general depth representations.
• Model Size Variants - Offers a multi-scale model hierarchy with various parameter counts to balance inference speed and accuracy.

Star history