Stream-based Active Learning

Temporal Predicted Loss for Stream-based Active Learning in Perception

⏱️ British Machine Vision Conference (BMVC) 2023

Overview

In the real world, data doesn’t sit in a neat pool waiting to be selected — it streams in continuously. Traditional pool-based active learning can’t keep up. You need to decide right now whether a sample is worth labeling.

We introduce Temporal Predicted Loss (TPL), a stream-based active learning method that exploits the temporal structure inherent in perception data. By leveraging the sequential nature of sensor streams, TPL makes real-time labeling decisions without ever needing to store or revisit large data pools.

Key Contributions

  • 🌊 True stream-based AL — operates on live data streams without requiring a stored pool
  • ⏱️ Temporal awareness — exploits sequential properties of perception data for smarter selection
  • Real-time decisions — instant, on-the-fly labeling choices as data arrives
  • 🚗 Built for autonomy — designed for continuous data collection in autonomous driving pipelines

Why It Matters

Autonomous systems generate data at a relentless pace. TPL enables these systems to learn efficiently from streaming data — selecting only the most informative frames in real time and drastically reducing annotation costs.

(Schmidt & Günnemann, 2023)

References

2023

  1. Stream-based Active Learning by Exploiting Temporal Properties in Perception with Temporal Predicted Loss
    Sebastian Schmidt and Stephan Günnemann
    In Proceedings of the British Machine Vision Conference (BMVC), Oct 2023