Team ERIS: IWDD 2027 illegal waste dumping detection
Weights of Team ERIS's entry to the International Contest on Illegal Waste Dumping Detection (IWDD 2027). A frozen ImageNet-21k ViT-B/16 with bottleneck adapters and a temporal transformer head classifies 3.2 s windows (16 frames at 5 fps) as normal or dumping; a threshold decoder turns the window probabilities into one onset per video.
iwdd_s4.pth holds {'model': state_dict} for the VideoModel in the submission's model.py;
test.py downloads it from this repository when no local copy is given.
Paper and code
This model is the Window Classifier of the paper "Parameter-Efficient Image-to-Video Adaptation for Action Spotting of
Illegal Waste Dumping" (WACV 2027 Workshops, WasteVision). The code is at
https://github.com/arkel23/PEI2V_WasteVision27; its README section "Pretrained weights" shows how to download this file
and score the public-test windows with tools.train --test_only.
Citation
@inproceedings{rios2027pei2v,
title={Parameter-Efficient Image-to-Video Adaptation for Action Spotting of Illegal Waste Dumping},
author={Rios, Edwin Arkel and Hu, Min-Chun and Lai, Bo-Cheng},
booktitle={IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)},
year={2027}
}
@inproceedings{greco2027evolution,
title={The Evolution of Illegal Waste Dumping Detection in the Age of Foundation Models},
author={Greco, Antonio and Ricciardi, Andrea Vincenzo and Sansone, Carlo and Vento, Bruno},
booktitle={IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)},
year={2027}
}