Abstract
CrossDet introduces a novel anchor-free object detection framework that represents objects using "cross lines" rather than traditional rectangular boxes or discrete points. By adaptively growing horizontal and vertical lines to fit an object's boundaries, the model captures continuous internal features while filtering out background noise, leading to superior accuracy on PASCAL VOC and MS-COCO benchmarks.
Introduction
In the world of computer vision, object representation is the foundation of detection. For years, two paradigms have dominated: anchor-based methods using predefined rectangular boxes and point-based methods using center points or keypoints. Anchor-based methods often include background "noise" in the corners and require tedious manual tuning. Point-based methods, while avoiding anchor-tuning, are discrete and often miss the continuous appearance information inside an object, making it difficult to determine exact boundaries in crowded or overlapping scenes.
CrossDet proposes a middle ground: a representation that is continuous enough to capture the object's "body" but sparse enough to ignore background interference.
Methodology
The core innovation is the Crossline Representation. Instead of a box, an object is represented by a horizontal line and a vertical line that intersect. These lines "grow" during training to match the object's extent.
Key Components
- Crossline Extraction Module (CEM): This module uses "axis-aware pooling" to gather context from the entire row and column, then selectively samples features along the predicted cross lines. This ensures the model "sees" the object's internal structure.
- Decoupled Regression Mechanism: Standard detectors try to predict width, height, and center all at once. CrossDet splits the task: the horizontal branch handles horizontal offsets and width, while the vertical branch handles the vertical axis. This reduces the mathematical optimization space, making it easier for the network to converge.
- Multi-stage Refinement: The architecture uses a multi-stage pipeline where coarse cross lines are predicted first and then refined into precise boundaries in subsequent steps.
Key Insights
- Lines over Boxes: Rectangular boxes are inherently noisy because objects are rarely perfect rectangles. Cross lines act as a "skeleton" that captures discriminative features (center and edges) without picking up background clutter in the corners.
- Continuous vs. Discrete: Unlike point-based methods, cross lines provide a continuous path of features. This helps the model perceive the change in features as it hits an object's boundary.
Experimental Results
The researchers evaluated CrossDet on two gold-standard benchmarks: PASCAL VOC and MS-COCO. CrossDet consistently outperformed both anchor-based and point-based predecessors:
- PASCAL VOC: Using a ResNet-50 backbone, CrossDet achieved 50.9% mAP, outperforming ATSS (49.3%) and Faster R-CNN (46.5%).
- MS-COCO: On the more challenging COCO dataset, CrossDet reached 48.4% AP (using ResNet-101 with Deformable Convolutions), surpassing other high-end detectors like BorderDet (47.2%) and RepPoints (45.0%).
- Ablation Studies: The authors showed that the CEM was vital; removing it caused a significant drop in performance (~2.5% AP), proving that feature sampling along lines is as critical as the lines themselves.
Limitations and Future Work
While effective, CrossDet relies on axis-aligned horizontal and vertical lines, which may be less optimal for heavily rotated objects (e.g., in aerial imagery). Additionally, the multi-stage nature and CEM module add architectural complexity compared to "one-shot" detectors. Future work could involve more complex line combinations or applying this representation to instance segmentation tasks.
Summary updated