Abstract
Anchor-based detectors are widely used in object detection. To improve the accuracy of object detection, multiple anchor boxes are intensively placed on the input image, yet most of which are invalid. Although the anchor-free method can reduce the number of useless anchor boxes, the invalid ones still occupy a high proportion. On this basis, this paper proposes a multi-scale center point object detection method based on parallel network to further reduce the number of useless anchor boxes. This study adopts the parallel network architecture of hourglass-104 and darknet-53 of which the first one outputs heatmaps to generate the center point for object feature location on the output attribute feature map of darknet-53. Combining feature pyramid and CIOU loss function this algorithm is trained and tested on MSCOCO dataset, increasing the detection rate of target location and the accuracy rate of small object detection. Though resembling the state-of-the-art two-stage detectors in overall object detection accuracy speed.
Introduction
Traditional "anchor-based" detectors, such as Faster R-CNN or YOLOv3, rely on densely placing thousands of predefined bounding boxes (anchors) across an image. This approach suffers from three major flaws: extreme class imbalance where over 90% of anchors are background, significant computational waste from calculating Intersection over Union (IoU) for tens of thousands of boxes, and high sensitivity to manually tuned hyperparameters like anchor size and aspect ratio.
While anchor-free methods like FCOS have emerged to address these issues, they often struggle with high negative sample rates or imprecise boundary regression. This research introduces a multiscale detection pipeline that eliminates anchors entirely, focusing on center-point detection to improve both efficiency and precision.
Methods: Parallel Network Architecture
The core innovation is a Parallel Network Architecture that decouples localization from attribute regression. The system splits the workload between two specialized backbones:
- Hourglass-104: This network generates multiscale heatmaps to identify the precise center point of objects.
- Darknet-53: This backbone extracts attribute features. Instead of scanning the whole image, the system uses the center points found by the Hourglass network to index directly into the Darknet feature maps to regress the object's width, height, and boundaries.
Key technical components include Feature Pyramid Networks (FPN) for handling objects of varying sizes and the CIoU (Complete IoU) Loss function. CIoU improves accuracy by accounting for the distance between center points and the consistency of aspect ratios. Notably, the model removes the "center-ness" branch used in models like FCOS, as the heatmap-based approach directly identifies high-quality center points, reducing computational overhead.
Key Results
The proposed method, designated as Ours_CIoU, was evaluated on the challenging MSCOCO dataset. It demonstrated strong performance, particularly in balancing speed and accuracy compared to existing state-of-the-art models.
As shown in the results table, the proposed method achieved an Average Precision (AP) of 46.0, significantly outperforming YOLOv3 (33.0) and FCOS (41.0). A major highlight is the performance on small objects, where it reached an APS of 30.1, surpassing most one-stage detectors. While the inference time (182ms) is slower than YOLOv3, it remains significantly faster than high-accuracy two-stage detectors like TridentNet (1429ms), offering a superior balance for high-precision applications.
Conclusion
By decoupling "where" an object is from "what" its dimensions are, the parallel network architecture reduces redundant computations and improves localization. The heatmap-based indexing effectively slashes the number of invalid cells processed by the model. While the Hourglass-104 backbone introduces some computational cost, the gains in accuracy—especially for small objects—mark a significant advancement in anchor-free object detection. Future work will focus on optimizing the backbone for real-time performance.
Summary updated