Husper

Husper is a sensing-based personalized adaptive bitrate (ABR) framework for video streaming that integrates real-time user behavior signals with network metrics to optimize user quality of experience (QoE). This repository provides an end-to-end pipeline covering raw video preparation, data processing, model training, and Unity-based system integration.

Implementation Overview

• Video Preparation: Segment raw videos into DASH-compatible chunks at multiple bitrates using video_preparation/phase4.py, segmenter.py, and quality_segmenter.py.
• Model Training: In husper/, define reward, CQL training scripts, and Jupyter notebooks for reproducible experiments.
• Playback Integration: Unity project in video_player[unity]/ that queries ONNX-exported policies at runtime to select bitrates.

Folder Structure

• video_player_unity/
  Unity scenes, C# wrappers, and sensor-capture modules for real-time VR playback under Husper.
• husper/
  Core algorithm code: biosignal feature extraction, reward design, offline CQL training, and policy export.
• video_preparation/
  Scripts to generate key-frame videos, segment into fixed‐length chunks, transcode to multiple quality tiers, and produce DASH manifests.