Derive per-pixel height uncertainty from the adaptive-bin distribution, propagate it through tiled inference and TTA, and publish ndsm_std_m.npy for the DepthWizard frontend.
Enhance GPU processing and batching capabilities; update README for clarity on test-time augmentation and batch sizes
- Implemented DPT trunk architecture in models/dpt.py for feature extraction.
- Added exponential moving average functionality in models/ema.py to stabilize training.
- Created DINOv3 encoder in models/encoder.py for robust block finding and layer-wise learning rate decay.
- Developed heads for regression and classification tasks in models/heads.py, including gated fusion mechanism.
- Introduced custom loss functions in models/losses.py to address specific training challenges.
- Implemented test-time augmentation strategies in models/tta.py to enhance model predictions.
- Updated requirements.txt to include necessary dependencies for the new model components.
