aquatax4

AquaTax-4 is a leakage-audited benchmark for detecting floating macro-debris in freshwater imagery. It pools 2,592 images and 5,825 bounding boxes in four material classes (plastic 5,538, paper 123, metal 118, glass 46) from three public sources: AquaTrash: 369 images, 469 boxes FloW-Img: 2,000 images, 5,272 boxes AquaSurf-Malnad-223: 223 images, 84 boxes on 26 images; the other 197 are verified debris-free and kept as background The corpus is split 70/15/15 into 1,814 training, 389 validation and 389 test images (seed 20260729). No images are redistributed. FloW-Img is distributed by its authors for non-commercial research use only. The release instead contains everything needed to rebuild the exact corpus from your own copies of the three sources: a per-image manifest (split, source, original file path, SHA-256 hash, box counts per class); split files for training, validation and test; all bounding boxes of the two CC BY 4.0 sources (AquaTrash, AquaSurf-Malnad-223); rebuild_corpus.py, which finds every image by its SHA-256 and regenerates the FloW-Img boxes from FloW-Img's own annotation files; the perceptual-hash leakage-audit script, and the evaluation and bootstrap scripts; per-image predictions of all five detectors, so every confidence interval and paired comparison in the article can be recomputed without a GPU; the DINO-Swin-B training configurations, with field-by-field diffs of the ablation; MEASURED_RESULTS.md, listing every number reported in the article. Sources. Obtain them from: AquaTrash: https://doi.org/10.1016/j.cscee.2020.100026 FloW-Img: https://doi.org/10.1109/ICCV48922.2021.01077 (on request from its authors) AquaSurf-Malnad-223: https://doi.org/10.5281/zenodo.22945111 Code repository: https://github.com/shaziyasherin/aquatax Licences: code under MIT; manifest, splits, annotations and results under CC BY 4.0. Source images remain under their own licences.