DPPD: Deformable Polar Polygon Object Detection

Yang Zheng, Oles Andrienko, Yonglei Zhao, Minwoo Park, Trung PhamWorkshops @ CVPR 2023 · DOI · arXiv

Introduction

In the field of computer vision, a persistent trade-off exists between computational efficiency and descriptive detail. Standard object detection frameworks typically utilize rectangular bounding boxes, which are computationally inexpensive but fail to accurately represent the geometry of complex, non-rectangular objects such as diagonal crosswalks or irregular traffic signs. Conversely, instance segmentation provides pixel-level masks but often requires significant computational resources, making it unsuitable for real-time deployment on edge devices like those used in autonomous driving.

DPPD (Deformable Polar Polygon Object Detection) bridges this gap by representing objects as flexible polygons within a polar coordinate system. Unlike previous polar-based methods that use fixed-angle rays, DPPD allows vertices to move freely in both radius and angle. This "deformable" approach enables the model to achieve high-fidelity shape alignment while maintaining the real-time speeds necessary for safety-critical applications.

Methodology

The core innovation of DPPD is its Deformable Polar Representation. Rather than predicting a radius for a predefined set of fixed angles (e.g., every 10 degrees), the network predicts a sparse set of $k$ vertices. Each vertex is defined as a pair of $(radius, angle)$, allowing the model to concentrate vertices at high-curvature points and corners where they are most needed.

The DPPD architecture overview, highlighting the classification and regression heads that output deformable polar coordinates.

Key Technical Components

  • Angle Decoding: The model predicts "angle deltas" between successive vertices. These are processed via a cumulative sum and normalized to ensure the resulting polygon closes perfectly within a $0$ to $2\pi$ range.
  • Differentiable Resampling: To facilitate training, the sparse predicted vertices and the ground truth are resampled into a dense set of $m$ raypoints ($m > k$). This operation is fully differentiable, allowing the network to learn fine-grained shape details during the training phase while remaining sparse and efficient during inference.
  • Polygon Regression Losses: The framework employs a composite loss function consisting of an Origin Loss (to stabilize the polar center), a Polar IoU Loss (to maximize shape overlap), and a Smoothness Loss (to prevent jagged edges by penalizing high-frequency oscillations).

Key Insights and Convergence

A primary finding of this research is that flexibility is more valuable than density. A small number of flexible vertices (e.g., 12 or 36) can outperform a much larger number of fixed-angle rays because the model naturally learns to place vertices at critical geometric landmarks. Furthermore, the authors found that using the geometric centroid as the polar origin provides significantly more stability than the center of a bounding box, particularly when dealing with partially occluded objects.

Evolution of the predicted polygon over training iterations, demonstrating how the deformable vertices align with the ground truth shape.

Experimental Results

The researchers evaluated DPPD on a massive in-house NVIDIA dataset (focused on traffic signs and crosswalks) and the public Cityscapes benchmark.

  • Accuracy: In crosswalk detection, DPPD-36 (using 36 vertices) achieved an F1-score of 61.05, significantly outperforming PolarMask-64's 54.08, despite the latter using nearly double the vertices.
  • Speed: DPPD is highly optimized for edge hardware. On an NVIDIA DRIVE Orin chip, the polygon head adds negligible latency, with total inference taking approximately 2.68ms, compared to 3.24ms for PolarMask.
  • Cityscapes: DPPD outperformed Poly-YOLO, achieving an Average Precision (AP) of 11.96 (vs. 8.7) while maintaining a high frame rate of 56.1 FPS on a Tesla V100 GPU.

Conclusion and Future Work

DPPD demonstrates that deformable polar polygons offer a compelling alternative to both bounding boxes and pixel masks for real-time object detection. While the current model assumes star-convex shapes (where a single center point can reach all boundary points), it provides a robust foundation for high-fidelity perception. Future work may involve extending this approach to 3D Bird’s-Eye-View (BEV) mapping and exploring more advanced transformer-based backbones to further enhance detection accuracy.

Summary updated

BibTeX
@inproceedings{zheng2023dppd,
  author = {Yang Zheng and Oles Andrienko and Yonglei Zhao and Minwoo Park and Trung Pham},
  title = {{DPPD: Deformable Polar Polygon Object Detection}},
  booktitle = {Workshops @ CVPR},
  year = {2023},
  doi = {10.1109/cvprw59228.2023.00013},
}
Figure