Abstract
Corner-guided object detection offers precise boundary localization by identifying top-left and bottom-right corners, mimicking human-like bounding box construction. However, these detectors often struggle with the "grouping problem" in crowded or repetitive scenes, where pairing the correct corners becomes ambiguous. This paper introduces Corner Affinity, a robust grouping algorithm that integrates geometric structure and semantic context. By leveraging a dual-factor approach, Corner Affinity significantly improves detection accuracy, achieving a 35.8% AP boost on the Citypersons dataset and revitalizing the corner-guided detection paradigm.
The Grouping Challenge in Corner-Guided Detection
While most modern detectors are center-guided—predicting an object's center and regressing its dimensions—corner-guided detection is inherently more constrained and potentially more precise. Corners have fewer degrees of freedom (satisfying only two boundaries) compared to centers (which must satisfy four).
Despite these advantages, detectors like the original CornerNet face a significant hurdle: the grouping problem. In scenes with multiple similar objects, such as a row of pedestrians or identical aircraft, the model often fails to determine which top-left corner belongs to which bottom-right corner. Previous heuristic methods frequently produced unreliable pairs and messy bounding boxes when objects overlapped.
Methods: The Corner Affinity Algorithm
The authors propose replacing simple embedding distances with Corner Affinity, a dual-factor mechanism that couples geometric logic with semantic similarity.
1. Structure Affinity (SA)
Structure Affinity captures "shallow construction knowledge." For every detected corner, the network regresses the expected width and height of the object. A valid pair is identified if the boxes formed by individual corner predictions exhibit a high Intersection-over-Union (IoU). This is further refined by "corner drifting," a bias term measuring how accurately predicted vectors point toward the opposite corner.
2. Contexts Affinity (CA)
Contexts Affinity captures "high-level semantic knowledge" using associative embeddings. A "pull-push" loss ensures that corners belonging to the same object share similar feature values, while corners from distinct objects are pushed apart in the feature space.
3. The Fusion Function
These two affinities are fused using a Gaussian-like function: $$CornerAffinity = \max(SA \cdot \exp(-\frac{CA^2}{\sigma}), 0)$$ This design ensures that a grouping is only accepted if it satisfies both geometric logic (SA) and semantic similarity (CA). In crowded scenes, the semantic distance (CA) acts as a decay factor for the structural score, effectively filtering out false positives.
Experimental Results
The Corner Affinity model was evaluated across three diverse benchmarks:
- MS-COCO: Achieved 46.3% AP, a 5.8% improvement over the CornerNet baseline, outperforming some Transformer-based models like DETR.
- Citypersons: Demonstrated a staggering 35.8% AP boost (from 29.1% to 64.9%), proving its resilience against heavy occlusion in pedestrian detection.
- UCAS-AOD: Improved by 17.2% AP, showing high reliability in aerial imagery where identical objects are often arranged in regular patterns.
Conclusion and Future Work
The "pseudo-3D" perspective of Corner Affinity—where SA provides the 2D base geometry and CA provides the "depth" to separate overlapping objects—proves that corner-guided detection remains a powerful alternative to center-based methods. While current performance is somewhat dependent on the underlying backbone's heatmap quality, the robust grouping logic introduced here opens new avenues for bionic detection strategies and potential applications in 3D object detection.
Summary updated