Detecting Tear Gas Canisters With Limited Training Data

Abstract

Human rights investigations often entail triaging large volumes of open source images and video in order to find moments that are relevant to a given investigation and warrant further inspection. Searching for instances of tear gas usage online manually is laborious and time-consuming. In this paper, we study various object detection models for their potential use in the discovery and identification of tear gas canisters for human rights monitors. CNN based object detection typically requires large volumes of training data, and prior to our work, an appropriate dataset of tear gas canisters did not exist. We benchmark methods for training object detectors using limited labelled data: we fine-tune different object detection models on the limited labelled data and compare performance to a few shot detector and augmentation strategies using synthetic data. We provide a dataset for evaluating and training tear gas canister detectors and indicate how such detectors can be deployed in real-world contexts for investigating human rights violations. Our experiments show that various techniques can improve results, including fine-tuning state of the art detectors, using few shot detectors, and including synthetic data as part of the training set.