monocular depth estimation
**Monocular depth estimation** is the **prediction of dense depth maps from a single RGB image using geometric cues learned from data** - despite no explicit stereo baseline at inference, models infer relative distance from perspective, texture, and semantic priors.
**What Is Monocular Depth Estimation?**
- **Definition**: Map each pixel to an estimated depth value from one camera frame.
- **Inference Constraint**: Single-image input without direct triangulation.
- **Output Type**: Relative depth or metric depth depending on training setup.
- **Model Families**: CNN encoders, transformer decoders, and hybrid geometry-aware networks.
**Why Monocular Depth Matters**
- **Hardware Simplicity**: Depth perception without dedicated depth sensors.
- **Wide Applicability**: Useful in AR, robotics, autonomous driving, and scene understanding.
- **Data Availability**: Can leverage large image datasets and self-supervised video training.
- **Pipeline Foundation**: Supports obstacle reasoning and 3D reconstruction tasks.
- **Cost Efficiency**: Enables scalable depth deployment on commodity cameras.
**Depth Cues Used by Models**
**Perspective and Geometry**:
- Vanishing points and converging lines imply depth structure.
**Semantic Priors**:
- Known object sizes and scene context guide distance estimation.
**Texture and Blur Patterns**:
- Gradient density and focus cues correlate with depth.
**How It Works**
**Step 1**:
- Encode RGB image into multi-scale feature hierarchy capturing local and global context.
**Step 2**:
- Decode features into dense depth map with scale-aware refinement and optional uncertainty prediction.
Monocular depth estimation is **a high-impact perception capability that extracts 3D structure from ordinary camera imagery** - strong models combine learned semantics with geometric consistency for reliable depth predictions.