AdelaiDet

AdelaiDet implements anchor-free instance-level object detection using the Fully Convolutional One-Stage (FCOS) detector on top of Detectron2 to perform per-pixel prediction for precise instance recognition in bioinformatics and medical imaging.

Key Features:

• Integration with Detectron2: Integrates with Detectron2 to leverage its object detection architecture for instance-level recognition.
• Anchor-Free Detection (FCOS): Implements the Fully Convolutional One-Stage (FCOS) detector that predicts object locations without predefined anchor boxes.
• Anchor-based method comparison: Operates without anchors used by RetinaNet, SSD, YOLOv3, and Faster R-CNN, avoiding anchor-box management and related IoU computations during training.
• Per-pixel dense prediction: Uses per-pixel dense predictions aligned with dense prediction tasks such as semantic segmentation for granular detection.
• Post-processing with NMS: Relies on non-maximum suppression (NMS) for post-processing detections.
• Reduced hyper-parameter sensitivity: Eliminates anchor-related hyper-parameters to reduce sensitivity associated with anchor-box tuning.
• Improved detection accuracy: Demonstrates improved detection accuracy by removing anchor-box complexities and hyper-parameter sensitivities.

Scientific Applications:

• Bioinformatics research: Instance-level recognition for identifying specific structures or patterns within biological imaging and related data.
• Computer vision and image analysis: Accurate object detection and instance recognition in image-analysis projects without anchor-based constraints.
• Medical imaging and downstream applications: Precise instance detection applicable to medical imaging and other downstream scientific analyses requiring high localization accuracy.

Methodology:

Built on Detectron2, AdelaiDet uses the FCOS anchor-free framework to predict object locations on a per-pixel basis, aligns with dense prediction approaches such as semantic segmentation, omits anchor boxes (reducing anchor-related hyper-parameters and IoU computations during training), and applies non-maximum suppression (NMS) for post-processing.