Explicit and Implicit Box Equivariance Learning for Weakly-Supervised Rotated Object Detection

Rotated object detection has emerged as a preferred alternative to the conventional Horizontal Box (HBox) object detection in various tasks. However, the majority of existing detection datasets are still annotated using HBox. Re-annotating these datasets for rotated object detection is notably inefficient and labor-intensive. Recently, weakly-supervised detectors based directly on HBox have provided a solution to bridge the gap between the rotated bounding box (RBox) prediction and the weak annotations. Nonetheless, they tend to overlook the potential impact of factors such as bounding box position, aspect ratio, and scale on angle prediction. Additionally, inconsistencies arise when evaluating rotation variations, hindering the inference performance. In light of these limitations, we propose an explicit and implicit equivariance detector (EIE-Det) to predict fine-grained RBox based on HBox annotations. Along with a shared backbone and neck structure, our proposed EIE-Det comprises a weakly supervised (WS) branch for predicting coarse RBox from HBox labels, an explicit equivariance (EE) branch for learning rotation consistency, and an implicit equivariance (IE) branch for learning position, aspect ratio, and scale consistency. Furthermore, we introduce the Tanimoto Coefficient for vector-based bounding box regression, enabling a comprehensive assessment of model prediction bias, thereby eliminating interference from variations in box aspect ratio and scale. Experiments demonstrate that our method not only significantly surpasses state-of-the-art models supervised with HBox but also achieves accuracy comparable to fully-supervised detectors (75.74
$\%$
and 58.50
$\%$

$AP_{50}$
on DOTA and DIOR-R, respectively).