The Australian Capital Territory Branch of the Statistical Society will be holding its next meeting on Tuesday 30th June in person and online via Zoom. For those attending in person, there will be the opportunity to meet for dinner afterwards in Canberra City.
Date: Tuesday 30th June 2026
Time: 5:45pm (AEST) followed by dinner (details below)
Venue: Allan Barton Forum, Level 2, ANU College of Business & Economics, Building 26C, The Australian National University (ANU Building 26C)
Online by Zoom: https://anu.zoom.us/j/82610896950?pwd=7zbH8nCJsyX8ZBVdqQPi8AVxTlSYML.1
Meeting ID: 826 1089 6950
Password: 939291
The zoom link will be open from 5.30pm. RSVP is not required.
Title: An End-to-End Video-Based System for Street Food Safety Inspection Using Vision-Language Models and Multi-Source Truth Discovery
Ensuring food safety in street food environments remains a major public health challenge in developing countries, where traditional inspection methods rely on in-person officer assessments that are costly, difficult to scale, and often subject to individual bias. Such evaluations can vary due to contextual familiarity and subjective judgement, leading to inconsistent assessment of critical hygiene infrastructure, including handwashing stations, dishwashing areas, waste disposal systems, and water storage facilities.
This talk presents an end-to-end video-based system for automating street food safety inspection using vision-language models (VLLMs). The system replaces direct officer observation with a scalable pipeline that operates on short recorded videos of vendor environments. It extracts representative visual evidence from video data and uses a lightweight VLLM to generate structured assessments of food safety compliance.
On top of this system, the work focuses on improving the reliability of generated assessments through multi-source supervision. This includes annotations from field officers, crowdworkers under different visual conditions, and multiple VLLM-based systems. Truth discovery methods are applied to aggregate these heterogeneous and noisy signals into more consistent estimates of underlying ground truth labels. These aggregated labels are then used to refine and retrain vision-language models, improving robustness under real-world annotation noise and domain variability.
Presenter: Dr Patrick (Weihao) Li
Dr Patrick (Weihao) Li is a Postdoctoral Research Fellow at the Research School of Finance, Actuarial Studies and Statistics, Australian National University. He holds a Bachelor of Commerce (majoring in Business Analytics), an Honours degree in Econometrics, and a PhD in Statistics from the Department of Econometrics and Business Statistics at Monash University. His research focuses on machine learning, data visualization, visual inference, and computer vision, with a particular interest in automating the evaluation of diagnostic graphics and understanding how people interpret them.
Dinner: After the talk, there will be a dinner at 7.15pm at SoLita Pizzeria, Restaurant & Bar, 143 London Circuit, Canberra
If you plan to attend the dinner, please RSVP by 5:00 pm on Monday 29 June by entering your details in the SSA ACT Branch dinner attendance sheet, or by contacting Warren Muller at wjmstats@outlook.com or 0407 916 868. Please treat your RSVP as a firm commitment rather than an expression of interest.
If you need to withdraw after the deadline, please remove your name from the attendance sheet and notify Warren by phone or text on 0407 916 868. Please note that last-minute requests to attend during the event cannot be accommodated, so be sure to RSVP through the attendance sheet in advance.
NOTE: SSA ACT branch are offering discounts to SSA early career and student members who attend dinner! For this meeting, dinners will be a fixed charge of $15 for student members and $25 for early career members.
@StatSocAus
