K-Fold Ensemble of Hiera and VideoMAE for Illegal Waste Dumping Detection

초록

Illegal waste dumping is an important environmental and public-health problem, and fixed CCTV cameras provide a scalable sensing modality for continuous monitoring. The IWDD Contest 2026 (WASTEVISION @ WACV 2026) focuses on video-level dumping detection and timestamp localization of the dumping moment. We present a multi-model solution that fine-tunes two complementary pretrained video backbones, Hiera and VideoMAE V2, with a multi-task head for binary classification and normalized timestamp regression. To reduce overfitting under limited labeled videos, we train each backbone with 5-fold cross validation and build a cross-fold ensemble (5 Hiera + 5 VideoMAE). We then conduct a lightweight grid search (44 settings) over the classification threshold and inter-architecture ensemble weights. On the provided labeled development set (400 videos), the best configuration (Hiera: VideoMAE =0.8: 0.2, threshold τ=0.5) achieves a composite F1-score of 91.99%, timestamp accuracy of 92.0% (within the contest tolerance window, δ s in [-3,+10] s), and 1.77 s mean absolute error for timestamp prediction.