Active Learning for Lane Detection: A Knowledge Distillation Approach

Fengchao Peng, Chao Wang, Jianzhuang Liu, Zhen YangICCV 2021 · OpenReview

Abstract

Lane detection is a key task for autonomous driving vehicles. Currently, lane detection relies on a huge amount of annotated images, which is a heavy burden. Active learning has been proposed to reduce annotation in many computer vision tasks, but no effort has been made for lane detection. Through experiments, we find that existing active learning methods perform poorly for lane detection, and the reasons are twofold. On one hand, most methods evaluate data uncertainties based on entropy, which is undesirable in lane detection because it encourages to select images with very few lanes or even no lane at all. On the other hand, existing methods are not aware of the noise of lane annotations, which is caused by heavy occlusion and unclear lane marks. In this paper, we build a novel knowledge distillation framework and evaluate the uncertainty of images based on the knowledge learnt by the student model. We show that the proposed uncertainty metric overcomes the above two problems. To reduce data redundancy, we explore the influence sets of image samples, and propose a new diversity metric for data selection. Finally we incorporate the uncertainty and diversity metrics, and develop a greedy algorithm for data selection. The experiments show that our method achieves new state-of-the-art on the lane detection benchmarks. In addition, we extend this method to common 2D object detection and the results show that it is also effective.

Introduction

Lane detection is essential for autonomous driving, yet training robust deep models requires massive quantities of precisely annotated data. Active Learning (AL) offers a solution by automatically selecting the most informative unlabeled images for human annotation. However, traditional AL methods fail in lane detection for two primary reasons: the "Entropy Trap," where models prioritize images with no lanes at all, and a lack of awareness regarding noisy annotations caused by occlusion or poor visibility.

Lane detection task: identifying lane boundaries in complex urban environments.

Methodology

This paper introduces a novel framework that leverages Knowledge Distillation (KD) to identify informative samples. The architecture utilizes three models: a large Teacher ($M_T$), a small Student ($M_S$), and a Distilled Student ($M_{S-KD}$) trained to mimic the teacher.

The proposed active learning framework using knowledge distillation and influence sets for data selection.

Uncertainty via Prediction Gaps

Instead of relying on entropy, the framework measures "prediction gaps" between these models:

  • $D_{ST}$: The gap between the Distilled Student and the Teacher. A large gap suggests the image contains knowledge that is difficult for a smaller model to learn.
  • $D_{SS}$: The gap between the two Students. Since noise is often random, a large gap between identical architectures (one distilled, one not) indicates likely annotation noise or heavy occlusion.

Diversity via Influence Sets

To prevent data redundancy, such as selecting multiple identical frames from a single video sequence, the authors utilize "Influence Sets" based on Reverse Nearest Neighbors (RNN). This metric identifies how many other unlabeled images a candidate image can effectively represent. A greedy algorithm then selects a subset that maximizes both uncertainty and global influence.

Experimental Results

The method was evaluated on major benchmarks including CULane, LLAMAS, and BDD100K. Key findings include:

  • State-of-the-Art Performance: The proposed method consistently outperformed baselines like Random selection, Entropy, and advanced methods like BADGE.
  • Efficiency: On the CULane dataset, the method achieved a higher F1-score with 30k images than competing methods achieved with 50k images.
  • Versatility: The framework is model-agnostic and successfully improved performance in 2D object detection tasks on BDD100K.

Conclusion

By using Knowledge Distillation as a diagnostic tool, this framework effectively filters out noisy data and "empty" images, focusing human annotation efforts on the most informative samples. While it requires more initial computation than simple entropy methods, the significant reduction in required training data makes it a powerful tool for scaling autonomous driving systems.

Summary updated

Abstract

Lane detection is a key task for autonomous driving vehicles. Currently, lane detection relies on a huge amount of annotated images, which is a heavy burden. Active learning has been proposed to reduce annotation in many computer vision tasks, but no effort has been made for lane detection. Through experiments, we find that existing active learning methods perform poorly for lane detection, and the reasons are twofold. On one hand, most methods evaluate data uncertainties based on entropy, which is undesirable in lane detection because it encourages to select images with very few lanes or even no lane at all. On the other hand, existing methods are not aware of the noise of lane annotations, which is caused by heavy occlusion and unclear lane marks. In this paper, we build a novel knowledge distillation framework and evaluate the uncertainty of images based on the knowledge learnt by the student model. We show that the proposed uncertainty metric overcomes the above two problems. To reduce data redundancy, we explore the influence sets of image samples, and propose a new diversity metric for data selection. Finally we incorporate the uncertainty and diversity metrics, and develop a greedy algorithm for data selection. The experiments show that our method achieves new state-of-the-art on the lane detection benchmarks. In addition, we extend this method to common 2D object detection and the results show that it is also effective.
BibTeX
@inproceedings{peng2021active,
  author = {Fengchao Peng and Chao Wang and Jianzhuang Liu and Zhen Yang},
  title = {{Active Learning for Lane Detection: A Knowledge Distillation Approach}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2021},
  url = {https://openreview.net/forum?id=-AFCbVTSzdS},
}
Figure