Reproduce “Towards Rotation Invariance in Object Detection” on OpenResearch

About this paper

Agastya Kalra, Guy Stoppi, Bradley Brown, Rishav Agarwal, Achuta Kadambi

An improved method for generating axis-aligned bounding box labels during rotation augmentation addresses the pervasive flaw of the "Largest Box Method," which often degrades object detection performance. Proposed by Akasha Imaging and UCLA, the Ellipse Method combined with a Rotation Uncertainty Loss function achieves enhanced rotational invariance and boosts AP to 32.72 on COCO at 30° rotation, substantially outperforming the 18.47 AP from the prior approach.

Key insights

• The "Largest Box Method," a common practice for bounding box label generation during rotation augmentation, is fundamentally flawed and significantly degrades object detector performance by creating overly large and inaccurate labels.
• Modeling an object as an ellipse is an effective and principled way to approximate its shape for generating rotated axis-aligned bounding box labels, leading to higher Expected IoU and improved model performance.
• Acknowledging and mitigating the inherent uncertainty in bounding box labels at specific rotation angles (e.g., around 45 degrees) through an adaptive loss function further enhances rotational robustness and prevents overfitting to approximations.

Summary by alphaXiv.