Abstract
This paper presents a novel one-stage object detector designed to address the challenges of multi-scale object detection and computational efficiency. The proposed architecture features a Shared Encoder-Decoder with Attention Mechanism (SEDAM) to extract multi-level features efficiently. A key innovation is the inclusion of a semantic-related branch that revises geometric position predictions using category-specific semantic information. Additionally, the authors demonstrate that minimum pooling is highly effective for detecting small objects. Evaluated on the MSCOCO 2014 benchmark, the method achieves a 3.8% AP improvement over the MNC baseline, demonstrating its effectiveness and efficiency.
Introduction
Object detection faces significant hurdles in handling objects across diverse scales while maintaining low computational overhead. Traditional Feature Pyramid Networks (FPNs) often rely on independent parameters for each scale, leading to increased memory and processing demands. Furthermore, standard anchor-free detectors primarily depend on geometric regression—predicting the distance from a pixel to bounding box boundaries—which can be brittle in complex scenes with varying lighting or occlusion.
As shown in Figure 1, while existing methods like SSD and FPN have paved the way for multi-scale detection, they often lack the adaptive capability to leverage semantic information for position refinement. This work introduces a "Pixel-Semantic Revising" approach that uses semantic features to "fix" geometric guesses, coupled with a shared encoder-decoder structure to optimize resource usage.
Methods
The researchers propose a three-pronged architectural innovation to improve detection accuracy and efficiency:
1. Shared Encoder-Decoder with Attention Mechanism (SEDAM)
Instead of using independent layers for every scale, the model employs a shared module. This "U-shape" structure consists of an encoder that down-samples features to capture broad context and a decoder that up-samples them to maintain spatial resolution. This sharing of parameters forces the model to learn more generalized representations across all object scales.
To further enhance feature extraction, the authors integrated a custom attention mechanism (Figure 6). This module fuses average-pooled and max-pooled features to highlight key information, using both channel and spatial attention to refine the feature maps before they are passed to the next convolutional layer.
2. Semantic-Related Center Branch
Standard anchor-free detectors (like FCOS) predict geometric distances. This paper adds a "semantic-revised" branch that uses the semantic features of a specific category to predict a "semantic center." This center acts as a corrective signal, adjusting and refining the initial geometric regression to be more robust to the actual scene content.
3. Minimum Pooling for Small Objects
In a departure from standard practices, the authors discovered that minimum pooling is more effective than max or average pooling for small object detection. While max pooling captures the strongest signal, minimum pooling helps preserve discriminative features that allow the model to better distinguish tiny objects from complex backgrounds.
Experimental Results
The model was trained and evaluated on the MSCOCO 2014 benchmark, using 82,783 images for training and 40,504 for validation. The proposed method, built on a ResNet-50 backbone, demonstrated significant performance gains:
- Overall Performance: Achieved a detection [email protected] of 49.8% on MSCOCO 2014.
- Baseline Comparison: It outperformed the MNC (Multi-task Network Cascades) baseline by 3.8% AP.
- Ablation Study: The semantic-revised module alone improved detection across all scales, with a notable 6.3% AP increase for large objects and a 1.3% increase for small objects compared to a baseline without the module.
- Efficiency: By sharing parameters in the SEDAM module, the detector maintains a lower computational footprint than complex multi-level structures like M2Det.
Conclusion and Future Work
The core insight of this research is that semantic information (what an object is) can effectively guide geometric localization (where an object is). By combining a shared encoder-decoder architecture with semantic position revision and specialized pooling strategies, the authors achieved a smarter, lighter detector.
Future work will explore integrating this semantic-revising approach into other backbone structures beyond ResNet and further refining the attention mechanism to handle extreme occlusion. Additionally, the surprising effectiveness of minimum pooling for small objects warrants deeper theoretical investigation to develop new specialized layers for tiny object detection.
Summary updated