Garberus: Three-Headed Video Classifier to Guard Against Illegal Waste Dumps

Detecting illegal waste dumping in videos presents a challenging multi-task problem, combining the need for binary classification of dump events, classification of the disposal type (static or dynamic), and regression of the exact timestamp of the dumping within the video. In addition, practical deployment requires an acceptable computation time on standard GPUs, enabling implementation on either embedded camera-integrated surveillance systems or remote host servers. To further support model robustness and reduce class imbalance, we introduce a new small, balanced dataset of web-scraped videos specifically curated for dump/no-dump detection. The proposed model is named Garberus, reflecting its three-headed architecture designed to tackle complementary sub-tasks. It follows a divide-andconquer strategy that decomposes the overall problem into three distinct binary tasks, each handled by a dedicated classification head. Leveraging a spatio-temporal feature extraction backbone such as TimeSformer, the model attaches specialized heads for dump detection, disposal-type classification, and timestamp prediction. This multi-head design effectively captures the complex temporal dynamics and content cues necessary for accurate detection and classification, while maintaining computational efficiency suitable for real-world surveillance applications.