Detecting Rotated Objects as Gaussian Distributions and its 3-D Generalization | JoVE Visualize

Area of Science:

Computer Vision
Machine Learning
Geometric Deep Learning

Background:

Traditional object detection methods use parameterized bounding boxes (BBox) and rotation angles, facing limitations in high-precision rotation detection.
Existing regression losses struggle to align with actual detection performance metrics, particularly at high Intersection over Union (IoU) thresholds.

Purpose of the Study:

To propose a novel approach for rotated object detection by modeling objects as Gaussian distributions.
To develop a new regression loss based on Gaussian distribution distances, such as Kullback-Leibler Divergence (KLD), for improved detection accuracy.
To extend the proposed method to 3-D object detection, including heading estimation.

Main Methods:

Modeling rotated objects as Gaussian distributions.
Utilizing Kullback-Leibler Divergence (KLD) as a regression loss for comparing Gaussian distributions.
Implementing an efficient Gaussian metric-based label assignment strategy.
Extending the 2-D approach to 3-D object detection with tailored heading estimation.

Main Results:

The Gaussian distribution approach overcomes boundary discontinuity and the square-like problem inherent in BBox methods.
The KLD loss effectively aligns with detection performance metrics, outperforming existing methods, especially at high IoU (e.g., 0.75).
Analysis of gradients reveals interpretable physical meanings, explaining the method's effectiveness in high-precision detection.
Superior performance demonstrated across twelve diverse public datasets (2-D/3-D, aerial, text, face images) with various base detectors.

Conclusions:

The proposed Gaussian distribution modeling and KLD loss offer a more effective and robust solution for rotated object detection compared to traditional BBox methods.
The method shows significant improvements in high-precision detection scenarios and generalizes well across various domains and dimensions.
The dynamic updating of BBox parameters under the Gaussian-based KLD loss provides interpretable insights into the model's effectiveness.