Overlapping Box Suppression Algorithm for Window-Based Object Detection
[ 1 ] Wydział Automatyki, Robotyki i Elektrotechniki, Politechnika Poznańska | [ DW ] doktorant wdrożeniowy
[2.2] Automatyka, elektronika, elektrotechnika i technologie kosmiczne
2024
rozdział w monografii naukowej / referat
angielski
- deep learning
- object detection
- focus and detect methods
EN In this work, we introduce an Overlapping Box Suppression algorithm, which can be used as a building block of any window-based object detection system to inhibit false-positive detections caused by partially visible objects. Window-based approaches are often used for small object detection to make the objects easier to detect by general-purpose detectors. However, this group of methods is typically associated with a reduced detection quality for larger objects due to the intersection of objects with detection windows. The proposed algorithm significantly improves the APm by 25% compared to the NMS-based filtering method in our detection system trained on images from the 3D-ZeF20 dataset.
325 - 330
20