Why Weakly Supervised Learning
Fully-labeled 3D ground truth is expensive and slow to collect. Weakly supervised approaches let you train useful models with partial, noisy, or indirect labels instead.
Common Techniques
| Technique | Use Case |
|---|---|
| Pseudo-labeling | Bootstrapping from a small labeled set |
| Multi-view consistency | Leveraging multiple camera angles |
| Regression heads | Predicting continuous joint positions |
Practical Tips
- Start with a strong pretrained backbone (e.g., EfficientNet) before adding task-specific heads
- Validate on a small, fully-labeled holdout set even if training data is weakly labeled
- Track per-joint error, not just an aggregate metric — it reveals where the model actually struggles