Towards Rotation Invariance in Object Detection

TL;DR

This paper introduces a novel rotation augmentation method for object detection that improves rotational invariance by addressing label ambiguity, using a differentiable approximation and Rotation Uncertainty Loss, validated across multiple datasets.

Contribution

It proposes a simple, effective rotation augmentation technique with a differentiable label approximation and Rotation Uncertainty Loss, enhancing rotational invariance in object detection models.

Findings

Improved AP, AP50, AP75 across datasets

Outperforms Largest Box Method in rotation tasks

Applicable to both one-stage and two-stage detectors

Abstract

Rotation augmentations generally improve a model's invariance/equivariance to rotation - except in object detection. In object detection the shape is not known, therefore rotation creates a label ambiguity. We show that the de-facto method for bounding box label rotation, the Largest Box Method, creates very large labels, leading to poor performance and in many cases worse performance than using no rotation at all. We propose a new method of rotation augmentation that can be implemented in a few lines of code. First, we create a differentiable approximation of label accuracy and show that axis-aligning the bounding box around an ellipse is optimal. We then introduce Rotation Uncertainty (RU) Loss, allowing the model to adapt to the uncertainty of the labels. On five different datasets (including COCO, PascalVOC, and Transparent Object Bin Picking), this approach improves the rotational invariance of both one-stage and two-stage architectures when measured with AP, AP50, and AP75. The code is available at https://github.com/akasha-imaging/ICCV2021.