PyTAGIT: a scalable and interactive human-in-the-loop tool for fast image annotation

Large-scale image annotation remains a critical bottleneck in training high-performing machine learning models, particularly for multi-class classification tasks. Existing annotation tools often lack dedicated support for classification and fail to integrate human-in-the-loop strategies that balance automation with expert supervision. In this paper, we introduce PyTAGIT, an open-source annotation framework designed for fast, scalable, and interactive multi-class labeling. PyTAGIT combines AI-assisted classification with intuitive user interactions such as drag-and-drop assignment, t-SNE-based exploration, and iterative refinement through confidence thresholds. The tool operates efficiently even on mid-range laptops and supports datasets with thousands of samples and dozens of classes. Extensive experiments on seven diverse datasets demonstrate that PyTAGIT achieves the best annotation accuracy on six datasets and second-best on the remaining one, significantly outperforming both traditional manual tagging and vision-language baselines (CLIP, LLM+CLIP, BLIP-2). Furthermore, PyTAGIT consistently completes full dataset annotation within a one-hour time budget, drastically reducing annotation time compared to manual and automatic alternatives.