Abstract
This paper addresses the inherent limitations of conventional bottleneck structures (diminished instance discriminability due to overemphasis on batch statistics) and decoupled heads (computational redundancy) in object detection frameworks by proposing two novel modules: the Instance-Specific Bottleneck with full-channel global self-attention (ISB) and the Instance-Specific Asymmetric Decoupled Head (ISADH). The ISB module innovatively reconstructs feature maps to establish an efficient full-channel global attention mechanism through synergistic fusion of batch-statistical and instance-specific features. Complementing this, the ISADH module pioneers an asymmetric decoupled architecture enabling hierarchical multi-dimensional feature integration via dual-stream batch-instance representation fusion. Extensive experiments on the MS-COCO benchmark demonstrate that the coordinated deployment of ISB and ISADH in the YOLO-PRO framework achieves state-of-the-art performance across all computational scales. Specifically, YOLO-PRO surpasses YOLOv8 by 1.0-1.6% AP (N/S/M/L/X scales) and outperforms YOLO11 by 0.1-0.5% AP in critical N/M/L/X groups, while maintaining competitive computational efficiency. This work provides practical insights for developing high-precision detectors deployable on edge devices.
Introduction
The YOLO (You Only Look Once) series has long been a cornerstone of real-time object detection. However, standard architectures often struggle with two primary issues: an over-reliance on batch statistics and computational redundancy in prediction heads. Traditional backbones utilize Batch Normalization (BN) to stabilize training, but this can inadvertently suppress unique instance-specific features, making it difficult to distinguish between similar objects in complex scenes. Furthermore, the decoupled heads used for classification and regression are often computationally heavy, hindering performance on edge devices.
YOLO-PRO addresses these challenges by introducing the Instance-Specific Bottleneck (ISB) and the Instance-Specific Asymmetric Decoupled Head (ISADH). These innovations allow the model to prioritize individual image characteristics and optimize computational resource allocation.
Methods
Instance-Specific Bottleneck (ISB)
The ISB module is designed to amplify discriminative features unique to each image by incorporating Instance Normalization (IN) alongside traditional methods. A key feature of the ISB is its Full-Channel Global Self-Attention mechanism. Unlike standard Vision Transformers that split channels into groups, the ISB uses a feature patch-channel reconstructor. This component unfolds spatial image patches into the channel dimension, allowing the self-attention mechanism to capture global context with linear complexity relative to the input size.
Instance-Specific Asymmetric Decoupled Head (ISADH)
To improve efficiency at the prediction end, YOLO-PRO utilizes an asymmetric architecture. Recognizing that bounding box regression is inherently less complex than classification, the ISADH employs a lightweight 1×1 convolutional kernel for regression while maintaining a 3×3 kernel for classification. Additionally, a dual-stream fusion branch integrates instance-specific features with batch-level features, ensuring that fine-grained details are preserved during the final prediction stage.
Results
The performance of YOLO-PRO was evaluated on the industry-standard MS-COCO 2017 benchmark across five computational scales: Nano (N), Small (S), Medium (M), Large (L), and Extra-Large (X).
Quantitative Performance
YOLO-PRO demonstrates significant improvements over existing state-of-the-art models:
- Superiority over YOLOv8: YOLO-PRO achieved an AP increase of 1.0% to 1.6% across all model scales.
- Comparison with YOLO11: The model outperformed the recent YOLO11 by 0.1% to 0.5% AP in critical categories (N/M/L/X) while remaining competitive in inference speed.
- Efficiency Gains: The Large (L) scale YOLO-PRO delivers accuracy comparable to the Extra-Large (X) scale YOLOv8, but with significantly lower computational overhead.
Qualitative Analysis
Eigen-CAM visualizations indicate that YOLO-PRO focuses more precisely on object boundaries and exhibits less background noise compared to its predecessors. This precision is particularly evident in crowded scenes where objects overlap, as the model is better equipped to handle ambiguous features through its instance-specific focus.
Conclusion
YOLO-PRO successfully enhances instance-specific discriminability and reduces computational redundancy through the coordinated use of ISB and ISADH modules. While the Nano version shows a minor latency trade-off (approximately 0.06ms slower than YOLOv8-N on a T4 GPU), the overall accuracy gains and efficiency improvements make it a robust candidate for edge deployment. Future research will focus on further optimizing these modules for extremely constrained hardware and extending the framework to tasks such as instance segmentation or pose estimation.
Summary updated