Weighted boxes fusion: Ensembling boxes from different object detection models

Boosting Ship Detection in SAR Images With Complementary Pretraining Techniques

“…Next, we will describe the OSD pretraining technique based on aerial images and the OSM pretraining technique by common representation learning in detail. Finally, we will introduce how the WBF strategy [31] fuse the predictions of OSM-SSD and OSD-SSD.…”

“…Here we only introduce the critical architecture of these two detection benchmarks, and we refer to their original paper [7], [36] to see a more detailed introduction. It is noted that our OSD pretraining technique, OSM pretraining technique and the WBF strategy [31] can be easily applied to other state-ofthe-art SAR ship detection benchmarks. Faster R-CNN [7], as the behalf of the two-stage object detection methods, consists of three modules: feature embedding network as the backbone to extract high-level features from the original images, region proposal network (RPN) generating the ship proposals for preliminarily predicting the location of ships and box regression network such as Fast R-CNN [37] finishing the binary classification and bounding box regression.…”

“…In contrast, the OSD-SSD can help identify and locate ships on sea area because of more ships' annotation information from the aerial image dataset. Considering these complementary advantages, we employ the weighted boxes fusion (WBF) strategy [31] to fuse the predictions of OSD-SSD and OSM-SSD. The WBF strategy utilizes confidence scores of all predicted bounding boxes to construct the averaged boxes, including confidence scores and coordinate locations, leading to a SAR ship detector with better generalization ability.…”

“…1) Improve the feature learning ability of ships in SAR images by proposing the OSD pretraining technique, which transfers the characteristics of ships in earth observations to SAR images by fully taking advantage of ships' annotation information from a large-scale aerial image dataset [29]; 2) Enhance the general feature embedding of SAR images by proposing the OSM pretraining technique, which transfers plentiful texture features from optical images to SAR images by common representation learning via bridge neural networks (BNN) [30] on the optical-SAR matching task; 3) Explore the complementary characteristics of the OSD pretraining-based SAR ship detector (OSD-SSD) and OSM pretraining-based SAR ship detector (OSM-SSD) and thus propose to employ the WBF strategy [31] to fuse the predictions of the two detectors for further improving detection results; 4) Conduct various experiments on four SAR ship detection datasets [32], [33], [34], [35] and three representative CNN-based detection benchmarks [7], [36] to verify the effectiveness and complementarity of the OSD-SSD and OSM-SSD detectors, and the state-of-the-art performance of the combination of the two detectors.…”

Seleção De Retângulos Envolventes Em Arquiteturas Para Detecção De Objetos

“…O método Weighted boxes fusion (WBF) (Solovyev et al, 2019) foi desenvolvido para o uso de combinações de diferentes resultados de modelos, propondo um ajuste das coordenadas e da confiabilidade dos resultados gerados. Nesta dissertação será utilizada um combinador de retângulos envolventes sobrepostos no mesmo objeto para a seleção do resultado da arquitetura.…”