TricubeNet: 2D Kernel-Based Object Representation for Weakly-Occluded Oriented Object Detection

Beomyoung Kim, Janghyeon Lee, Sihaeng Lee, Doyeon Kim, Junmo Kim

Abstract

Oriented object detection (OOD) is a critical task in computer vision, particularly for aerial imagery and dense retail environments. Traditional methods often rely on complex bounding box regression, which suffers from angle discontinuity and computational inefficiency. We propose TricubeNet, an anchor-free detector that represents objects using rotatable 2D Tricube kernels within a heatmap. By treating detection as a localization task via visual kernels rather than coordinate regression, TricubeNet effectively bypasses the boundary problem and improves performance in cluttered scenes. Our architecture incorporates a Multi-Angle Convolution (MAC) module for orientation-invariant feature extraction and a cascade refinement strategy for sharp heatmap generation. Experimental results on DOTA and SKU110K-R datasets demonstrate that TricubeNet achieves state-of-the-art accuracy while maintaining high efficiency.

1. Introduction

Detecting oriented objects—such as ships in satellite imagery or products on a retail shelf—presents unique challenges compared to standard horizontal detection. Most existing frameworks utilize anchor-based regression, requiring millions of candidate boxes at various scales and orientations. This approach leads to several significant issues:

  1. Angle Discontinuity: Regressing angles (e.g., the jump from -90° to +90°) causes mathematical instabilities and loss fluctuations during training.
  2. Computational Complexity: Calculating the Intersection-over-Union (IoU) for millions of rotated boxes is computationally taxing.
  3. Dense Packing: In cluttered environments, horizontal bounding boxes overlap excessively, causing standard Non-Maximum Suppression (NMS) to erroneously discard valid detections.
Comparison of traditional regression-based methods (5-offsets and 8-offsets) against the proposed heatmap-based TricubeNet. The heatmap approach avoids the instability of regressing multiple coordinates for dense objects.

TricubeNet addresses these problems by moving away from regression entirely. Instead, it localizes objects through visual cues in a heatmap, representing each object as a rotatable kernel.

2. Methodology

2.1 The Tricube Kernel Representation

Standard heatmap-based detectors typically use circular Gaussian kernels. However, Gaussians are rotationally symmetric and do not capture the elongated nature of oriented objects well. TricubeNet introduces the 2D Tricube kernel, defined by the function:

$$K(x, y) = (1 - |x|^3)^\gamma \cdot (1 - |y|^3)^\gamma$$

This kernel creates a more rectangular distribution that can be scaled and rotated to match the ground truth bounding box precisely. This representation allows the model to learn the spatial extent and orientation of an object without explicit coordinate regression.

2.2 Network Architecture

The overall pipeline of TricubeNet consists of a backbone for feature extraction followed by a specialized refinement stage.

The TricubeNet architecture overview, illustrating the backbone, the Heatmap Cascade Refinement stage utilizing Multi-Angle Convolution (MAC) modules, and the final heatmap output.

Multi-Angle Convolution (MAC) Module

To enable the network to learn features independent of an object's orientation, we designed the MAC module. Instead of using complex steerable filters, the MAC module rotates the input feature map at several fixed angles (e.g., 0°, 30°, 45°, 60°), applies standard $3 \times 3$ convolutions, rotates the features back to the original orientation, and concatenates them. This allows the network to perceive objects from multiple perspectives simultaneously.

Heatmap Cascade Refinement

Rather than predicting the heatmap in a single step, TricubeNet employs a progressive refinement strategy. Features pass through a sequence of MAC modules and $1 \times 1$ convolutions, iteratively "cleaning up" the heatmap. This ensures sharp, well-separated kernels even in extremely cluttered scenes where objects are weakly occluded.

2.3 Specialized Loss Functions

To optimize training, we introduce two specialized losses:

  • Size-Weight Mask (SWM): Standard Mean Squared Error (MSE) loss often favors larger objects. SWM applies weights inversely proportional to object size, ensuring the network maintains high sensitivity to small objects.
  • False-Positive Example Mining (FPEM): This loss balances the contribution of foreground (object) and background pixels, effectively reducing "ghost" detections in complex backgrounds.

3. Experimental Results

TricubeNet was evaluated on four diverse benchmarks, showing competitive performance across different domains:

  • DOTA (Aerial Imagery): Reached 75.26 mAP, outperforming established two-stage anchor-based models like RoI Transformer and SCRDet.
  • SKU110K-R (Retail): Achieved 57.7 mAP, a significant improvement over CenterNet (54.7) and DRN (55.9), proving its robustness in dense environments.
  • Scene Text (MSRA-TD500 & ICDAR2015): Demonstrated strong localization capabilities for text lines at various orientations.

By bypassing complex rotated NMS and instead using simple Connected Component Labeling (CCL) with OpenCV’s minAreaRect, TricubeNet offers a more streamlined and efficient post-processing pipeline.

4. Conclusion

TricubeNet represents a shift in oriented object detection by replacing coordinate regression with a robust 2D kernel-based heatmap representation. By utilizing the Tricube kernel and the MAC module, the model avoids the pitfalls of angle discontinuity and handles dense packing with high precision. Future work will focus on addressing heavy occlusion where kernels may merge and integrating sub-pixel offset predictions to further enhance localization for extremely small objects.

Summary updated

Abstract

We present a novel approach for oriented object detection, named TricubeNet, which localizes oriented objects using visual cues (i.e., heatmap) instead of oriented box offsets regression. We represent each object as a 2D Tricube kernel and extract bounding boxes using simple image-processing algorithms. Our approach is able to (1) obtain well-arranged boxes from visual cues, (2) solve the angle discontinuity problem, and (3) can save computational complexity due to our anchor-free modeling. To further boost the performance, we propose some effective techniques for size-invariant loss, reducing false detections, extracting rotation-invariant features, and heatmap refinement. To demonstrate the effectiveness of our TricubeNet, we experiment on various tasks for weakly-occluded oriented object detection: detection in an aerial image, densely packed object image, and text image. The extensive experimental results show that our TricubeNet is quite effective for oriented object detection. Code is available at https://github.com/qjadud1994/TricubeNet.
BibTeX
@inproceedings{kim2022tricubenet,
  author = {Beomyoung Kim and Janghyeon Lee and Sihaeng Lee and Doyeon Kim and Junmo Kim},
  title = {{TricubeNet: 2D Kernel-Based Object Representation for Weakly-Occluded Oriented Object Detection}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2022},
  doi = {10.1109/wacv51458.2022.00348},
}
Figure