Abstract
MidNet is a novel object detector for oriented ships in aerial imagery that replaces traditional rotation-angle regression with a five-point keypoint representation. By predicting a center and four edge midpoints, MidNet avoids mathematical instabilities such as boundary discontinuity and high angular sensitivity. Combined with Symmetrical Deformable Convolution (SDC), MidNet achieves state-of-the-art results on major benchmarks including HRSC2016 and FGSD2021.
Introduction
Detecting ships from a bird's-eye view is notoriously difficult because ships appear at arbitrary orientations and often have very high aspect ratios. Most existing detectors rely on anchor-based methods or angular regression. Anchor-based methods require heavy manual hyper-parameter tuning, while angular regression suffers from boundary discontinuity—where tiny prediction errors near periodic boundaries (e.g., -90° vs 90°) result in massive loss values. Furthermore, for long ships, even a 1-degree error in angle prediction causes a drastic drop in Intersection over Union (IoU), leading to poor detection scores.
The MidNet Architecture
MidNet moves away from angles entirely, treating ship detection as a keypoint estimation task. The model predicts five keypoints: the center of the ship and the midpoints of its four edges.
Five-Point Encoding and SDC
Unlike corners, which can be ambiguous when rotated, midpoints always correspond to a specific quadrant relative to the center, leading to more stable feature learning. To enhance this, MidNet utilizes a Symmetrical Deformable Convolution (SDC) layer. Since ships are structurally symmetrical, SDC allows a midpoint on one side of the ship to learn from the features of its symmetric counterpart, which is particularly effective for detecting ships with long wakes or partial obstructions.
Adaptive Grouping and Refinement
To link these points, the model predicts a "matching radius" for the center and "centripetal shifts" for the midpoints. If a midpoint's shifted position falls within a center's radius, they are grouped. Finally, an analytical geometry algorithm refines these points to construct a precise rectangular bounding box, enforcing geometric constraints as a powerful inductive bias.
Key Insights
Midpoints over Corners
Midpoints are more likely to fall directly on the object than corners. This ensures that the features extracted at midpoint locations are richer and more representative of the ship's actual structure.
Quadrant Consistency
No matter how much a ship rotates, the "top" midpoint stays in the top quadrant relative to the center. This eliminates the orientation definition ambiguity that often plagues corner-based detectors.
Experimental Results
The authors evaluated MidNet on three prominent datasets:
- HRSC2016: Achieved 90.52% AP, outperforming state-of-the-art models like Oriented R-CNN.
- FGSD2021: Achieved 86.50% mAP, a significant 1.04% improvement over previous best methods in fine-grained classification.
- DOTA: Achieved a competitive 86.42% AP for the ship category.
Ablation studies confirmed the design choices: switching from corners to midpoints improved accuracy by 2.23%, while the addition of SDC contributed an additional 0.87%.
Conclusion
MidNet demonstrates that an anchor-and-angle-free approach can overcome the fundamental flaws of angular regression in ship detection. While highly effective for rectangular vessels, future work will focus on adapting the model for extremely irregular shapes and exploring weakly supervised learning to reduce annotation requirements.
Summary updated