AiWo at IEEE BigData 2026: Listening before looking to cut the cost of video AI
We’re delighted to share that a paper by AiWo’s postdoctoral researcher Masoud Jalayer has been accepted to the Industrial and Government Track of the 2026 IEEE International Conference on Big Data (IEEE BigData 2026), a leading international forum for big data research, development and applications.
- Conference: https://bigdataieee.org/BigData2026/
- Date: 14 Dec 2026 – 17 Dec 2026, Phoenix, Arizona, USA
About the research
Hours-long recordings from wearable cameras are increasingly valuable for progress monitoring, quality control and safety in logistics, construction and manufacturing. Analyzing them with a vision–language model (VLM) is expensive, however, because the model is called on many short windows and the cost grows with every call. Existing methods for choosing which windows deserve a call rely on visual features, which means decoding the very video the budget is meant to save. This paper proposes audio-first triage: the windows are selected from the soundtrack alone, before any video frame is decoded. The approach is evaluated on EPIC-KITCHENS-100 and Ego4D.
Accepted paper and key contributions
- Pre-Decoding Acoustic Triage for Budgeted Vision–Language Captioning of Untrimmed Egocentric Video
Rather than training a conventional sound-event detector, the paper trains a lightweight audio gate to fire once per action, which matches what a call budget actually pays for. With the same frozen AudioSet-pretrained features, this change of objective improves action coverage by 4.0–10.8 percentage points at every call rate, without any domain-specific sound labels. A simple minimum-separation rule then spreads the calls over time. On EPIC-KITCHENS-100, this cuts 9–20% of VLM calls at matched coverage and beats uniform sampling on 22 of 24 held-out recordings. It also outperforms two recent visual keyframe selectors. Audio screening costs about 3% of a full VLM scan, while optical-flow screening costs about 75%.
Why it matters
Industrial video archives are long, but compute budgets are not. Because the triage runs entirely on audio, it can be combined with standard VLM efficiency techniques such as token compression or quantization. The paper also gives practitioners a simple way to estimate, before deployment, how much triage can save on their own recordings. Code and evaluation scripts are publicly available at https://github.com/masjalayer/PreDecoding-AcousticTriage.
Authors
Masoud Jalayer, Changyi Li, Yu Xiao (Aalto University)