HAISTA-NET addresses the limitations of fully automated instance segmentation algorithms by introducing human-assisted segmentation. The model utilizes human attention maps to enhance predictions for high-curvature, complex, and small-scale objects. By combining automated and interactive segmentation approaches, HAISTA-NET achieves superior results compared to state-of-the-art models like Mask R-CNN and Mask2Former. The Partial Sketch Object Boundaries (PSOB) dataset contains hand-drawn partial object boundaries representing object curvatures. HAISTA-NET architecture integrates human attention maps during training and inference, demonstrating improved performance in mask precision for challenging objects. A user-friendly interface allows users to interact with objects through partial strokes, enhancing annotation efficiency and model accuracy.
HAISTA-NET
Stats
HAISTA-NET outperforms Mask R-CNN, Strong Mask R-CNN, and Mask2Former with increases of +36.7, +29.6, and +26.5 points in APMask metrics.
PSOB dataset includes annotations from 30 users for 18,677 objects of different scales and curvature sections.
HAISTA-NET achieves a +26.5-point increase in APMask versus Mask2Former on the COCO dataset.
Quotes
"Practitioners typically resort to fully manual annotation, which can be a laborious process."
"Our human-assisted segmentation model augments existing networks to incorporate human-specified partial boundaries."
"We propose a novel approach to enable more precise predictions and generate higher-quality segmentation masks."