Abstract
In this paper, we make an effort to improve the accuracy of small and medium object detections of SSD (Single Shot Multibox Detector). To this end, we introduce a deconvolutional region magnification procedure in which the existing layers in SSD play a role in the region proposal network and the proposed regions are magnified for recognition. Moreover, features are also extracted from a shallow layer and a new feature pyramid is constructed on top of these structures. Then, features are contacted and fed into classification and regression modules as in SSD. The weights of the present model are obtained via a pre-training-re-training strategy. By evaluating the model performance on a test set assembled by the samples in the PASCAL VOC and MS COCO datasets, the present model shows that the mAPs (mean average precisions) of small and medium object detections are 42.4% and 74.7% respectively, which are 27.1% and 15.6% better than SSD. This proves the effectiveness of our proposed method.
The Small Object Problem in SSD
In the world of computer vision, the Single Shot Multibox Detector (SSD) is a legend for its speed. Unlike two-stage detectors (like Faster R-CNN) that first propose regions and then look at them closely, SSD does everything in one go. However, SSD has a persistent "small object problem."
To detect small things, SSD relies on early, "shallow" layers of the neural network. While these layers have high resolution, they lack semantic information—they see edges and textures but don't "understand" the object as well as deeper layers do. Conversely, by the time an image reaches the deep layers, small objects have often been blurred or downsampled into oblivion. The authors set out to bridge this gap: providing high-resolution details and rich semantic understanding simultaneously.
Methods: Magnifying the Feature Space
The authors propose a hybrid architecture that adds three major components to the classic SSD pipeline:
- Deconvolutional Region Magnification: The model uses the initial predictions of the SSD layers as "region proposals." It then applies a deconvolutional layer (
Deconv2) to these regions. This mathematically "magnifies" the feature maps, increasing their resolution so the network can see finer details of small objects. - Shallow Feature Integration: Standard SSD starts its detection pyramid at the
conv4_3layer. This paper goes even shallower, extracting features fromconv3_3. This ensures that the "raw material" for small objects isn't lost before the detection process begins. - New Feature Pyramid: On top of these magnified and shallow features, the authors construct a new feature pyramid. These features are concatenated and fed into classification and regression modules, similar to the original SSD pipeline.
- Pre-training/Re-training Strategy: To ensure stability, the model uses a two-stage training approach, first training on a general dataset and then undergoing a specific re-training phase to fine-tune the new magnification components.
Key Insights: The "Zoom-In" Intuition
The most powerful idea here is treating deconvolution as a "magnifying glass." By upscaling the feature maps of predicted regions, the model can perform classification on a higher-resolution representation of a small object than was previously possible in a single-shot framework.
As shown in the visualization above, the model analyzes feature maps from various stages (Conv4_3, Fc7, Conv8_2) to refine its understanding of the scene. This allows the model to correctly identify "covered" or "shaded" objects that the original SSD might completely ignore or misclassify.
Results and Performance
The improvements over the baseline SSD are significant, particularly for the target object sizes:
- Small Object Accuracy: The mean Average Precision (mAP) reached 42.4%, a massive 27.1% improvement over standard SSD.
- Medium Object Accuracy: The mAP reached 74.7%, a 15.6% improvement over standard SSD.
- Speed: The model runs at 24 FPS (frames per second) on an Nvidia Titan Xp. While this is slower than the original SSD (55 FPS), it sits right at the threshold of real-time performance (25 FPS) while offering much higher reliability.
Conclusion
By borrowing the "Region Proposal" logic from slower, more accurate models and baking it into the SSD architecture using efficient convolutional operations, the authors have demonstrated that you don't have to choose between high resolution and high semantics. This magnification procedure effectively "injects" semantic meaning into high-resolution maps, making it a viable solution for applications where small object detection is critical.
Summary updated