Multi-Scale Receptive Field Detection Network

Haoren Cui, Zhihua WeiIEEE Access 2019 · DOI

Abstract

Deep convolutional neural networks have significantly advanced object detection, yet scale variation remains a primary challenge. We introduce the Multi-Scale Receptive Field Detection Network (MS-RFDN), a one-stage detector designed to handle objects across diverse scales. By integrating a novel Scale-Specific Concatenation (SSC) module, the network merges features from dense and dilated blocks to optimize receptive fields for specific scales. Our approach demonstrates substantial improvements in small object detection, achieving competitive performance on the PASCAL VOC 2007 benchmark.

Introduction: The Challenge of Scale

In computer vision, detecting objects that vary significantly in size—from a tiny bird in the distance to a large vehicle in the foreground—is a persistent hurdle. Standard CNNs often possess a fixed receptive field that may not be ideal for all object sizes simultaneously. While techniques like image pyramids or feature pyramids offer some relief, they often come with high computational costs or insufficient spatial context. MS-RFDN addresses this by tailoring the network's receptive field to the scale of the target objects.

Proposed Method: MS-RFDN Architecture

The MS-RFDN architecture is built upon three primary structural innovations designed for efficiency and multi-scale robustness.

1. Refined DenseNet Backbone

The network utilizes a modified DenseNet-121 as its foundation. To preserve fine-grained spatial details necessary for small object detection, the initial $7 \times 7$ convolution is replaced with three successive $3 \times 3$ convolutions. This modification ensures a more gradual increase in the receptive field, retaining critical edge information often lost in standard backbones.

2. Scale-Specific Concatenation (SSC) Module

The core of MS-RFDN is the SSC module. This component processes feature maps through two parallel paths: a standard dense block and a dilated block. The dilated convolutions are specifically configured so that their output matches the receptive field size of the dense block. By concatenating these two distinct feature sets, the network captures both high-resolution details and broader spatial context for the same image region.

Architecture of the Multi-Scale Receptive Field Detection Network (MS-RFDN) featuring the Scale-Specific Concatenation Module (SSCM).

3. Multi-Scale Detection Strategy

Following a one-stage detection logic, the network performs predictions at three different downsampling levels (8x, 16x, and 32x). This multi-level approach allows different layers to specialize in detecting objects of varying sizes, further enhancing the model's versatility.

Experimental Results

The MS-RFDN was evaluated on the PASCAL VOC 2007 test set after training on the combined VOC 2007 and 2012 datasets.

  • Detection Accuracy: The model achieved a mean Average Precision (mAP) of 78.72%, outperforming popular one-stage detectors like YOLOv2 (76.75%) and SSD512 (78.60%). It sits just behind the heavier YOLOv3 (79.12%).
  • Small Object Performance: The introduction of the SSC module led to a performance gain of approximately 0.8% to 1.0%, with a significant boost in recall for smaller instances.
  • Efficiency: With a prediction time of 104ms, MS-RFDN maintains a competitive balance between high precision and operational speed, though it is slightly slower than YOLOv3 due to the increased parameter count of the SSC modules.

Conclusion

MS-RFDN successfully addresses the scale variation problem through targeted receptive field engineering. By combining dense and dilated features within the SSC module, the network achieves state-of-the-art accuracy for a one-stage detector. Future research will focus on model compression techniques to reduce the computational overhead introduced by the SSC modules, aiming for even faster real-time performance on complex datasets like MS COCO.

Summary updated

Abstract

Deep convolutional neural networks have contributed much to various computer vision problems including object detection. However, there are still many problems to be solved. Scale variation across object instances is one of the major challenges for object detection. In this paper, we propose a multi-scale receptive field detection network (MS-RFDN), a one-stage approach to detect objects of different scales in the image. The proposed network combines predictions of different scales from feature maps of different scales and receptive fields. To generate s scale-specific feature maps in specific layer, we design a scale-specific concatenation module (SSC module). This scale-specific feature maps are merged from the dense block and dilated block, which has the same size of the receptive field. Through our multi-scale layer network structure and scale-specific feature maps, our model has a significant improvement in small object detection. On the VOC 2007 test dataset, our method almost achieves the effect of the state-of-the-art one-stage methods, which confirmed the effectiveness of our model.
BibTeX
@inproceedings{cui2019multi,
  author = {Haoren Cui and Zhihua Wei},
  title = {{Multi-Scale Receptive Field Detection Network}},
  booktitle = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2942077},
}
Figure