self-supervised monocular depth
**Self-supervised monocular depth estimation** is the **training approach that learns single-image depth prediction from unlabeled stereo pairs or monocular video using photometric consistency losses** - it delivers practical depth models without dense ground-truth annotations.
**What Is Self-Supervised Monocular Depth?**
- **Definition**: Depth network trained by reconstructing one view from another through predicted disparity or depth.
- **Training Data**: Stereo pairs, monocular sequences, or mixed setups.
- **Inference Mode**: Single-image depth prediction at deployment.
- **Scale Behavior**: Metric scale can be learned with known stereo baseline during training.
**Why It Matters**
- **Annotation-Free Learning**: Eliminates need for expensive lidar-labeled datasets.
- **Scalable Pretraining**: Uses large video corpora across domains.
- **Practical Deployment**: Produces lightweight monocular depth predictors for edge devices.
- **Domain Transfer**: Easier adaptation through self-supervised fine-tuning.
- **SLAM Support**: Provides useful priors for visual odometry and mapping.
**Training Recipe**
**Photometric Reconstruction**:
- Warp source view to target using predicted depth and relative pose.
- Minimize pixel and structural similarity losses.
**Regularization Terms**:
- Edge-aware smoothness to reduce depth noise.
- Occlusion masking and auto-masking for dynamic regions.
**Multi-Scale Supervision**:
- Depth outputs at several scales improve optimization stability.
- Coarse-to-fine refinement improves detail.
**How It Works**
**Step 1**:
- Predict depth map and relative pose from unlabeled image pairs or sequences.
**Step 2**:
- Reconstruct target view through differentiable warping and optimize consistency losses.
Self-supervised monocular depth is **a scalable route to practical depth perception that learns from geometric consistency rather than manual labels** - it is a core technique for modern low-cost 3D vision systems.