Bagging R-CNN: Ensemble for Object Detection in Complex Traffic Scenes

Pengteng Li, Ying He, Dongfu Yin, F. Richard Yu, Pinhao Song

Abstract

Generic object detection methods have achieved preferable results, but it is still challenging to detect objects from complicated traffic scenes like extreme illumination and adverse weather. The existing methods are not robust enough to be extended to new complex traffic scenes. To address this issue, we leverage the idea of ensemble learning for strong robustness and propose a novel Bagging R-CNN framework. Specially, we design a bagging classification branch that uses adaptive sampling to train base learners and make them different from each other. The final predictions are the ensemble of the base learners, achieving strong robustness to the challenging objects. For localizing more accurately, a progressive regression branch is proposed in which bounding boxes are continuously optimized for high quality. Extensive experiment results on TJU-DHD-traffic and Pascal VOC datasets show that our Bagging R-CNN achieves superior detection accuracy over state-of-the-art methods. The source code can be found at https://github.com/PungTeng/BaggingRCNN.

Introduction

Standard object detection models often struggle when moved from clean, curated datasets to the chaotic environment of real-world traffic. The authors identify three primary hurdles:

  1. Extreme Illumination: Blinding sunlight or deep shadows can obscure features.
  2. Adverse Weather: Rain, fog, and snow create "vague" objects that are hard to distinguish from the background.
  3. Data Distribution Imbalance: Training sets are often dominated by "easy-to-learn" samples, leading to models with high variance that perform poorly on difficult, complex cases.

While some researchers try to fix this with image enhancement (like dehazing or brightening), these modules are often fragile. Bagging R-CNN instead looks to ensemble learning to build inherent robustness into the model architecture itself.

Main challenges in traffic environments: (a) extreme illumination, (b) diverse scenes, and (c) adverse weather.

Methodology

The Bagging R-CNN framework consists of two novel components integrated into a standard R-CNN pipeline:

1. Bagging Classification Branch

Instead of a single classification head, this branch uses multiple "base learners" (C1, C2, C3). To ensure these learners don't just learn the exact same things (homogenization), the authors use Adaptive Resampling. The weights of training samples are adjusted dynamically; if a sample was poorly classified by the previous base learner, its weight is increased for the next one. This forces subsequent learners to focus on the "hard" samples that their predecessors missed. The final classification score is the average of all base learners, leveraging "collective intelligence" for better generalization.

2. Progressive Regression Branch

In many detectors, classification and regression share the same features, which can lead to sub-optimal localization. Bagging R-CNN decouples these tasks. The regression branch consists of multiple stages (B1, B2, B3) that continuously refine the bounding box coordinates. Each stage takes the output of the previous one and optimizes it further, leading to much higher localization quality.

Overview of the proposed Bagging R-CNN framework, showing the parallel bagging classification and progressive regression branches.

Results and Discussion

The authors evaluated their model on two key benchmarks: TJU-DHD-traffic, a high-resolution dataset collected from driving cars, and Pascal VOC, a classic generic object detection benchmark.

Bagging R-CNN consistently outperformed state-of-the-art (SOTA) methods:

  • TJU-DHD-traffic: Using a Swin-T backbone, the model achieved 63.0% AP and 87.1% AP50, significantly higher than standard Cascade R-CNN or Dynamic R-CNN.
  • Pascal VOC: The model reached 81.4% mAP, a +1.4% improvement over the strong Cascade R-CNN baseline.

Qualitative improvements were notable in detecting small or low-contrast objects, such as cyclists in dark or rainy conditions, where baseline models often fail. Stability analysis also showed that the bagging approach helps prevent the model from overfitting on easy samples, maintaining consistent performance across training epochs.

Conclusion

Bagging R-CNN demonstrates that ensemble learning is a powerful tool for building robust object detectors capable of handling the "messy" reality of traffic scenes. By ensuring diversity among base learners and progressively refining localization, the framework achieves superior accuracy without relying on fragile preprocessing steps. Future work could involve exploring more efficient ways to maintain this diversity or applying the bagging concept to the feature extractor itself.

Summary updated

Abstract

Generic object detection methods have achieved preferable results, but it is still challenging to detect objects from complicated traffic scenes like extreme illumination and adverse weather. The existing methods are not robust enough to be extended to new complex traffic scenes. To address this issue, we leverage the idea of ensemble learning for strong robustness and propose a novel Bagging R-CNN framework. Specially, we design a bagging classification branch that uses adaptive sampling to train base learners and make them different from each other. The final predictions are the ensemble of the base learners, achieving strong robustness to the challenging objects. For localizing more accurately, a progressive regression branch is proposed in which bounding boxes are continuously optimized for high quality. Extensive experiment results on TJU-DHD-traffic and Pascal VOC datasets show that our Bagging R-CNN achieves superior detection accuracy over state-of-the-art methods. The source code can be found at https://github.com/PungTeng/BaggingRCNN.
BibTeX
@inproceedings{li2023bagging,
  author = {Pengteng Li and Ying He and Dongfu Yin and F. Richard Yu and Pinhao Song},
  title = {{Bagging R-CNN: Ensemble for Object Detection in Complex Traffic Scenes}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=ZGM44xfY3h},
}
Figure