Energy Efficient Pupil Tracking Based on Rule Distillation of Cascade Regression Forest

Sangwon Kim, Mira Jeong, ByoungChul Ko

TL;DR

This paper introduces a novel, energy-efficient pupil tracking algorithm called Cascade Deep Regression Forest (CCD-RF) with rule distillation. This method significantly reduces computational load and model size, making it ideal for on-device machine learning applications while maintaining high accuracy.

What Problem Does This Paper Address?

As human-computer interaction (HCI) becomes more sophisticated, accurate pupil tracking is crucial for applications ranging from virtual reality to advanced driver assistance systems. While deep neural networks (DNNs) have achieved state-of-the-art results, they are computationally expensive and operate as "black boxes," making them unsuitable for resource-constrained on-device systems. The challenge is to develop a pupil tracking system that is both accurate and efficient enough to run locally on devices, while also being interpretable.

Methods

The core of this work is a Cascade Deep Regression Forest (CCD-RF), which is a multi-layered regression forest structure designed for coarse-to-fine pupil estimation. Unlike traditional DNNs, RFs offer better interpretability and faster inference. To further enhance efficiency and interpretability, the authors introduce a rule distillation technique. This process involves:

  • Rule Extraction: Identifying and analyzing the decision rules within each regression tree of the CCD-RF.
  • Rule Contribution Analysis: Quantifying the importance of each rule in contributing to the overall prediction accuracy.
  • Rule Distillation: Eliminating low-importance rules based on a user-defined distillation rate. This simplifies the model by removing redundant or less impactful decision paths.

This distillation process aims to create a more transparent and computationally lighter model without sacrificing significant accuracy. Additionally, a consistency check between the left and right eyes is incorporated to improve robustness against false detections.

Datasets

The authors evaluated their proposed method using two publicly available datasets:

  • GI4E dataset: Contains 1236 color images from 103 subjects with 12 different gaze directions, varying illumination, facial poses, and occlusions.
  • BioID dataset: Consists of 1521 grayscale images from 23 subjects, known for its challenging variations in illumination, poses, and occlusions.

Key Insights and Ideas

  • Rule Distillation for Efficiency: The key innovation is using rule distillation to prune a regression forest. This is a more interpretable approach to model compression compared to typical DNN compression techniques, as it directly simplifies the decision-making logic.
  • Interpretability and Efficiency Synergy: By distilling rules, the model becomes more transparent (easier to understand how it works) and significantly more energy-efficient and computationally lighter, making it suitable for embedded systems.
  • Coarse-to-Fine Cascade: The multi-layer cascade structure allows for progressive refinement of pupil position estimation, starting with a coarse localization and moving towards finer details.
  • On-Device Viability: The significant reduction in parameters and operations makes this method a strong candidate for real-time pupil tracking on devices with limited processing power and battery life.

Key Results

  • Accuracy: The proposed method achieves comparable or superior accuracy to state-of-the-art methods, particularly at stricter error thresholds (e.g., $e_d \le 0.05$). For instance, on the GI4E dataset, the proposed method (with 10% rule removal) achieved 99.3% accuracy at $e_d \le 0.05$, while on BioID, it reached 93.4% accuracy at the same threshold.
  • Model Simplification: Rule distillation led to substantial reductions in parameters and operations. For example, removing 50% of rules on the GI4E dataset reduced parameters by 38% and operations by 5% while maintaining accuracy.
  • Speed: The method demonstrates a significant speedup. On a CPU, the proposed method (rule ratio 0.9) achieved a processing time of 40.2 ms, which is considerably faster than GPU-based CNN models.
  • Energy Efficiency: Comparative experiments showed that the proposed method consumed significantly less battery power than CNN-based models, achieving 1.5 times less consumption than MobileNetV3.

Limitations and Future Work

  • Rule Redundancy: The authors acknowledge that the distilled model may still contain some overlapped features, trivial rules, and unnecessary parameters, preventing it from being a complete "white-box" model.
  • Future Work: The authors suggest future research to focus on reducing rule redundancy through hierarchical rule importance sharing and feature correlation analysis. They also aim to further optimize the model for even greater efficiency and interpretability.

Summary updated

BibTeX
@inproceedings{kim2020energy,
  author = {Sangwon Kim and Mira Jeong and ByoungChul Ko},
  title = {{Energy Efficient Pupil Tracking Based on Rule Distillation of Cascade Regression Forest}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=k91OKABLjQ},
}

Loading related content...

Figure