Authors:
Bruno Padilha
and
João Eduardo Ferreira
Affiliation:
Institute of Mathematics and Statistics (IME-USP), University of São Paulo, São Paulo, Brazil
Keyword(s):
Active Learning, Out-of-Distribution Classification, Collaborative Image Classification, Big Data Labeling.
Abstract:
Streaming applications in video monitoring networks generate datasets that are continuously expanding in terms of data amount and sources. Thus, given the sheer amount of data in these scenarios, one big and fundamental challenge is how to reliably automate data annotation for downstream tasks such as object detection, image classification, object tracking among other functionalities. In this work, we propose a novel active learning strategy based on multi-model collaboration able to self-annotate training data, providing only a small initial subset of human verified labels, towards incremental model improvement and distribution shifts adaptation. To validate our approach, we collected approximately 50,000 hours of video data sourced from 193 security cameras from University of São Paulo Monitoring System (USP-EMS) during the years 2021-2023, totaling 7.3TB of raw data. For experimental purposes, this work is focused on identification of pedestrians, cyclists and motorcyclists result
ing in 3.5M unique objects labeled with accuracy between 92% to 96% for all cameras. Time-stamped data along with our incremental learning method also facilitate management of naturally occurring distribution shifts (e.g., weather conditions, time of the year, dirty lenses, out-of-focus cameras). We are currently working to release this dataset in compliance with local data privacy legislation.
(More)