Abstract
Traditional neural objection detection methods use multi-scale features that allow multiple detectors to perform detecting tasks independently and in parallel. At the same time, with the handling of the prior box, the algorithm's ability to deal with scale invariance is enhanced. However, too many prior boxes and independent detectors will increase the computational redundancy of the detection algorithm. In this study, we introduce Dubox, a new one-stage approach that detects the objects without prior box. Working with multi-scale features, the designed dual scale residual unit makes dual scale detectors no longer run independently. The second scale detector learns the residual of the first. Dubox has enhanced the capacity of heuristic-guided that can further enable the first scale detector to maximize the detection of small targets and the second to detect objects that cannot be identified by the first one. Besides, for each scale detector, with the new classification-regression progressive strapped loss makes our process not based on prior boxes. Integrating these strategies, our detection algorithm has achieved excellent performance in terms of speed and accuracy. Extensive experiments on the VOC, COCO object detection benchmark have confirmed the effectiveness of this algorithm.
Introduction
Most modern object detectors, such as SSD, YOLOv3, and Faster R-CNN, rely heavily on prior boxes (anchors)—pre-defined bounding boxes of various scales and aspect ratios tiled across the image. While effective, this approach introduces significant redundancy, manual heuristic dependency, and ignores the continuous nature of scale in the real world. Generating thousands of anchors creates massive computational overhead and severe class imbalance.
DuBox addresses these issues by moving away from anchors entirely. It introduces a cooperative multi-scale framework where higher-level detectors learn from the errors of lower-level ones, significantly reducing computational redundancy while maintaining high precision.
Methodology
DuBox introduces several novel components to achieve high-performance "no-prior box" detection:
Hook-based Regression
Instead of matching anchors, DuBox defines "hooks" (pixels on the feature map). Each hook directly regresses four offsets—the distance to the top, bottom, left, and right edges of the object. This eliminates the need for manually tuned anchor sizes.
Residual Dual Scale Detectors
This is the core architectural innovation. Instead of having two independent detection branches, the high-level detector learns the residual (the error) of the low-level detector. A "Bbox Bridge Module" passes information between scales, forcing them to work synergistically. This "division of labor" prevents scales from competing and allows the network to refine rough estimates into precise detections.
Classification-Regression Progressive Strapped (CRPS) Loss
To ensure that a hook only predicts a class if it can also accurately regress the box, the authors use an IoU Gate. If the predicted box's IoU with the ground truth is too low, the classification loss for that hook is ignored during training. This synchronizes the two tasks and prevents "high-confidence, low-accuracy" detections.
Experimental Results
DuBox was evaluated on standard industry benchmarks, including PASCAL VOC and MS COCO, demonstrating that anchors are not necessary for speed or accuracy.
- Accuracy: On VOC 2007, DuBox800 achieved 82.31% mAP, outperforming many two-stage and one-stage competitors like Faster R-CNN and SSD.
- Speed: The DuBox320 variant runs at 50 FPS on an NVIDIA Titan X while maintaining a high 79.31% mAP, making it a top-tier choice for real-time applications.
- COCO Performance: On the MS COCO dataset, DuBox800 reached 39.52% AP, significantly better than YOLOv3 and RetinaNet500.
Conclusion
DuBox demonstrates that a cooperative, residual-based approach to multi-scale detection can outperform traditional independent detector architectures. By aligning classification confidence with regression accuracy through the CRPS loss and utilizing a "no-prior box" strategy, DuBox achieves a state-of-the-art balance between speed and precision. Future work may explore extending this residual chain to more than two scales to further enhance performance.
Summary updated