active-learning-wildlife

What problem does it solve?

Active learning workflows for wildlife detection using the HILDA framework (v0.3.0). The core workflow minimizes human annotation effort by intelligently selecting which images an expert should review: Train → Predict → Select uncertain samples → Export to CVAT/Label Studio → Expert corrects → Download corrected annotations → Retrain → Repeat.

Core Features & Use Cases

• Supports hands-on active learning with CVAT and Label Studio integration for pre-annotations and corrected outputs.
• Implements a three-way sampling strategy family (RGB Contrast, Embedding Clustering, and Logit Uncertainty) to optimize data efficiency.
• Provides a modular loop orchestrator that handles training, prediction, sample selection, annotation export, and learning-curve logging.

Quick Start

Run the active-learning loop for wildlife detection using HILDA with your dataset.