Add egocentric rotation support to centroid cropping inference

[gitttt-1234]

Summary

Extends centroid cropping with optional orientation-based rotation during inference, similar to the egocentric rotation already available in the training pipeline.

What's Different from Standard Centroid Cropping?

Standard centroid cropping extracts fixed-size crops around detected centroids without rotation. This branch adds optional egocentric rotation that aligns instances to a canonical orientation before cropping:

• Images and keypoints are rotated to align orientation anchor points along the positive x-axis
• Crops are extracted from the rotated images
• Predictions are inverse-rotated back to the original orientation
• Coordinates are correctly transformed from crop space to full image space

Usage Example

from omegaconf import OmegaConf
from sleap_nn.training.model_trainer import ModelTrainer

if __name__ == "__main__":

    config = OmegaConf.load("config.yaml")
    trainer = ModelTrainer.get_model_trainer_from_config(config, orientation_anchor_parts=["nose", "top of neck", "left ear", "right ear"])
    trainer.train()

Key Changes

• Add orientation_anchor_parts parameter throughout inference pipeline
• Apply per-instance egocentric rotation in CentroidCrop._generate_crops()
• Store rotation angles and apply inverse rotation in FindInstancePeaks.forward()
• Fix coordinate transformation from crop to full image space

🤖 Generated with Claude Code