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. The extensive experimental results show that our TricubeNet is quite effective for oriented object detection.
Rethinking Oriented Object Detection
Detecting objects that aren't perfectly horizontal—like a plane in an aerial photo or a tilted book on a shelf—is surprisingly difficult for computers. Traditional oriented object detection (OOD) methods typically attempt to find objects by regressing five parameters: center $(x, y)$, width $(w)$, height $(h)$, and angle $( heta)$. This coordinate-based approach leads to several significant challenges:
- The Anchor Explosion: To cover every possible size and angle, models often need to evaluate millions of "anchor boxes" (up to 4 million for a single image), which is computationally expensive.
- Angle Discontinuity: Angles are mathematically tricky. In many coordinate systems, $-90^\circ$ and $+90^\circ$ represent the same orientation, but a neural network sees them as very different numbers. This causes "loss fluctuation" during training when an object's angle hovers near these boundaries.
- Dense Packing: In scenarios like retail shelves or satellite imagery, objects are packed tightly. Standard methods often accidentally delete valid detections during the cleaning phase (Non-Maximum Suppression) because their horizontal bounding boxes overlap too much.
TricubeNet offers a refreshing alternative: instead of guessing box coordinates and angles, it represents objects as specialized "heatmaps" that naturally encode orientation.
The TricubeNet Architecture
TricubeNet moves away from coordinate regression entirely, treating detection as a visual cue problem. By representing orientation as a shape in a heatmap rather than a number $( heta)$, the authors completely sidestep the angle discontinuity problem. The network simply learns to "draw" the object's footprint.
Key Components
- 2D Tricube Kernel: Instead of using a circular Gaussian blob (which has no orientation), the authors use a "Tricube" kernel $(1-|x|^3)^7$. This kernel provides a smooth gradient that helps the network learn the edges and orientation of an object simultaneously, preserving the object's aspect ratio directly in the heatmap.
- Multi-Angle Convolution (MAC): Standard convolutions struggle to recognize rotated features. The MAC module applies convolutions at four specific angles ($0, 30, 45, 60$ degrees) and merges the results, making the network rotation-invariant.
- Heatmap Cascade Refinement: The model uses a multi-stage process to progressively sharpen and refine the heatmap, ensuring higher precision in the final output.
- Size-Weight Mask (SWM): Standard loss functions often ignore tiny objects. SWM weights the loss inversely to the object's size, ensuring the model pays equal attention to small cars and large warehouses.
Experimental Results
TricubeNet achieves state-of-the-art or highly competitive results across several challenging benchmarks:
- DOTA (Aerial Imagery): Achieved 75.26 mAP, outperforming many complex two-stage anchor-based models like Gliding Vertex.
- SKU110K-R (Dense Retail): Reached 57.7 mAP, significantly better than other anchor-free methods like CenterNet (54.7) and DRN (55.9).
- Ablation Gains: The MAC module alone provided a 1.3% boost in accuracy, while the heatmap cascade added another 2.3%, proving that iterative refinement is crucial for heatmap-based detection.
Conclusion
By eliminating the need for millions of anchor boxes, TricubeNet becomes simpler to tune and more robust to different object densities. While it is excellent for "weakly-occluded" objects, future work may address heavy occlusion where objects almost entirely hide one another. Currently, the model provides a powerful, anchor-free alternative for high-precision oriented object detection.
Summary updated