Optimization for Arbitrary-Oriented Object Detection via Representation Invariance Loss

Abstract

Arbitrary-oriented objects exist widely in natural scenes, and thus the oriented object detection has received extensive attention in recent years. The mainstream rotation detectors use oriented bounding boxes (OBB) or quadrilateral bounding boxes (QBB) to represent the rotating objects. However, these methods suffer from the representation ambiguity for oriented object definition, which leads to suboptimal regression optimization and the inconsistency between the loss metric and the localization accuracy of the predictions. In this paper, we propose a Representation Invariance Loss (RIL) to optimize the bounding box regression for the rotating objects. Specifically, RIL treats multiple representations of an oriented object as multiple equivalent local minima, and hence transforms bounding box regression into an adaptive matching process with these local minima. Then, the Hungarian matching algorithm is adopted to obtain the optimal regression strategy. We also propose a normalized rotation loss to alleviate the weak correlation between different variables and their unbalanced loss contribution in OBB representation. Extensive experiments on remote sensing datasets and scene text datasets show that our method achieves consistent and substantial improvement. The source code and trained models are available at https://github.com/ming71/RIDet.

G-Rep: Gaussian Representation for Arbitrary-Oriented Object Detection

“…Additionally, some researchers have proposed using an eight-parameter QBB to represent the object and Smooth l 1 as regression loss to as the regression loss to regress the four corner points of QBB (Xu et al, 2020;Ming et al, 2021a;Liu et al, 2019;Feng et al, 2020). The recent anchor-free method CFA (Guo et al, 2021) uses the more flexible PointSet (i.e., point set) to represent the oriented object, inspired by the horizontal object detection method RepPoints (Yang et al, 2019b).…”

“…CSL (Yang & Yan, 2020) and DCL (Yang et al, 2021a) transform angular prediction from regression to classification. RIDet (Ming et al, 2021a) uses the representation invariant loss to optimize bounding box regression. Using the IoU value as the regression loss (Yu et al, 2016) has become a research topic of great interest in oriented object detection, and can avoid partial problems caused by the regression of the angle parameter or point ordering.…”

“…ii) As a special case of PointSet, QBB is defined as the four corners of a quadrilateral bounding box. In addition to the inherent problems of PointSet described above, QBB also suffers from the representation ambiguity problem (Ming et al, 2021a). Quadrilateral detection often sorts the points first (represented by green box Figure 1(b)) to facilitate the point matching between the ground-truth and prediction bounding boxes to calculate the final loss function.…”

“…To improve detection performance, many researchers have proposed solutions for some issues, which mainly include the boundary discontinuity (Yang et al, 2019a;Yang & Yan, 2020), square-like problem (Yang et al, 2021a;Qian et al, 2021), representation ambiguity (Ming et al, 2021a), and isolated points. Recent methods GWD (Yang et al, 2021c) and KLD (Yang et al, 2021d) have broken the paradigm of existing regular regression frameworks from a unified regression loss perspective.…”

Towards Weakly-Supervised Text Spotting using a Multi-Task Transformer

“…Our weakly-supervised training scheme is obtained by introducing a new loss function based on the Hungarian matching loss [4] that simultaneously optimizes the detection and recognition tasks. The Hungarian loss, which has shown promise in the field of object detection [4,9,29,44], is meaningful in our setting, where the matching explicitly uses the text content for the detection optimization. Our Hungarian loss, which we call Text Hungarian Loss, replaces the detection cost with a recognition cost in the matching criteria.…”