Receptive Field Block Net for Accurate and Fast Object Detection (1711.07767v3)

Abstract: Current top-performing object detectors depend on deep CNN backbones, such as ResNet-101 and Inception, benefiting from their powerful feature representations but suffering from high computational costs. Conversely, some lightweight model based detectors fulfil real time processing, while their accuracies are often criticized. In this paper, we explore an alternative to build a fast and accurate detector by strengthening lightweight features using a hand-crafted mechanism. Inspired by the structure of Receptive Fields (RFs) in human visual systems, we propose a novel RF Block (RFB) module, which takes the relationship between the size and eccentricity of RFs into account, to enhance the feature discriminability and robustness. We further assemble RFB to the top of SSD, constructing the RFB Net detector. To evaluate its effectiveness, experiments are conducted on two major benchmarks and the results show that RFB Net is able to reach the performance of advanced very deep detectors while keeping the real-time speed. Code is available at https://github.com/ruinmessi/RFBNet.

• The paper introduces the innovative RFB module, which mimics human visual receptive fields to enhance feature representation in SSD detectors.
• It achieves improved detection accuracy with mAP rising from 77.2% to 82.2% on Pascal VOC and competitive AP on MS COCO.
• The method balances computational efficiency and accuracy, enabling real-time detection at 83 FPS without significant overhead.

Receptive Field Block Net for Accurate and Fast Object Detection

The work presented in "Receptive Field Block Net for Accurate and Fast Object Detection" by Songtao Liu, Di Huang, and Yunhong Wang addresses a pivotal issue in object detection: balancing computational efficiency and detection accuracy. The research leverages the innovative concept of Receptive Field Block (RFB), inspired by the structural characteristics of receptive fields in the human visual system, to enhance feature representation in lightweight convolutional neural networks (CNNs), leading to both accurate and fast object detection.

Context and Motivation

Object detection methodologies have seen significant advancements with the advent of deep CNN architectures such as ResNet, Inception, and their derivatives. Despite their increased detection accuracy, these models suffer from substantial computational costs, making real-time applications infeasible. Single-stage detectors such as SSD and YOLO have shown promise in achieving real-time object detection; however, they often sacrifice detection accuracy, particularly on smaller objects or objects with complex contextual relationships.

Key Contributions

The RFB module, the core contribution of this paper, mimics the human visual cortex's receptive field characteristics by considering the relationship between the size and eccentricity of receptive fields. This module enhances feature discriminability and robustness without significantly increasing computational overhead. The researchers integrate the RFB module into the SSD framework, leading to the development of the RFB Net.

Design and Implementation of RFB

• Multi-Branch Convolution Layer: The RFB employs a multi-branch architecture where each branch consists of 1x1 convolutions followed by varying-sized n x n convolutions to simulate receptive fields of different sizes.
• Dilated Convolution Layers: These layers facilitate capturing spatial hierarchies and relationships at different scales, akin to the human visual system's eccentricity variations.

Empirical Validation

The effectiveness of the proposed RFB Net is demonstrated through extensive experiments on PASCAL VOC 2007 and MS COCO datasets. Key findings include:

1. Pascal VOC 2007: The RFB Net300 achieves an mAP of 80.5%, outperforming the standard SSD300 which achieves 77.2%. RFB Net512 further improves the detection performance to an mAP of 82.2%.
2. MS COCO: On the COCO dataset, RFB Net300 achieves an average precision of 30.3% at IoU 0.5: 0.95, significantly higher than SSD300. The larger RFB Net512 model achieves an average precision of 33.8%, competing closely with other state-of-the-art detectors while maintaining a high processing speed.

Comparative Analysis

The RFB-based detectors deliver competitive performance across various conditions:

• Accuracy vs. Speed: RFB Net300 runs at 83 FPS with an mAP of 80.5% on Pascal VOC 2007, making it one of the most efficient real-time detectors. On MS COCO, RFB Net512 achieves competitive accuracy (33.8% AP) at 30ms, outperforming many existing high frame rate detectors.
• Robustness and Discriminability: Thanks to the multi-branch convolution structure and dilated convolutions, the RFB module significantly enhances feature representation, leading to improved detection of small objects and robustness to spatial variations.

Theoretical and Practical Implications

The integration of biological principles into neural network architectures, as exemplified by the RFB module, opens avenues for further research at the intersection of computational neuroscience and deep learning. The practical implications are significant for applications requiring real-time object detection, such as autonomous driving, surveillance, and robotics.

Future Developments

Future research could explore:

• Integration with Advanced Backbones: Implementing the RFB module with more advanced lightweight backbones (e.g., MobileNet, ShuffleNet) could enhance both the computational efficiency and accuracy of object detectors.
• Adapting to Various Detection Frameworks: The adaptability of the RFB module to other one-stage and two-stage detection frameworks could be an area of extensive exploration.
• Further Exploration of Biological Mechanisms: Additional insights from the human visual system could inspire enhancements to CNN architectures, leading to even more robust and efficient models.

Conclusion

The RFB Net introduces a highly effective balance between accuracy and computational efficiency in object detection. By leveraging biologically inspired motifs, the authors successfully enhance feature representation in lightweight detectors, setting a new standard for real-time object detection. The quantitative improvements on standard benchmarks attest to the feasibility and potential of integrating neuroscience principles into deep learning methodologies.