depth refinement
**Depth refinement** is the **post-processing or learned correction stage that improves raw depth maps by sharpening boundaries, removing noise, and enforcing structural consistency** - it turns coarse predictions into geometry usable for high-precision tasks.
**What Is Depth Refinement?**
- **Definition**: Enhance initial depth outputs from sensors or networks using edge-aware filtering or learned residual correction.
- **Input Sources**: Monocular depth, stereo disparity, lidar completion, or fused depth.
- **Common Defects**: Edge bleeding, speckle noise, quantization, and hole artifacts.
- **Output Goal**: Cleaner depth with preserved discontinuities and stable surfaces.
**Why Depth Refinement Matters**
- **Boundary Accuracy**: Sharp depth edges are essential for segmentation and obstacle localization.
- **Surface Quality**: Reduced noise improves mesh reconstruction and mapping.
- **Temporal Stability**: Better refinement reduces flicker in video depth pipelines.
- **Planning Reliability**: Cleaner depth lowers false obstacle signals.
- **Visual Quality**: AR compositing and rendering depend on precise depth boundaries.
**Refinement Techniques**
**Guided Filtering**:
- Use RGB image edges to guide depth smoothing.
- Preserve discontinuities while denoising flat regions.
**Bilateral and Joint Bilateral Filters**:
- Weight smoothing by spatial and intensity similarity.
- Control cross-edge diffusion.
**Neural Refinement Heads**:
- Learn residual corrections from depth plus image context.
- Improve complex artifact cases beyond handcrafted filters.
**How It Works**
**Step 1**:
- Detect noisy and uncertain regions in initial depth map.
**Step 2**:
- Apply edge-aware filtering or learned residual correction and output refined depth.
Depth refinement is **the final quality-upgrade stage that makes raw depth estimates precise enough for reliable perception and interaction** - strong refinement preserves edges while suppressing spurious noise.