OliviaWuYJ/sc-floatwaste

FloatWaste

A multi-region benchmark for detecting floating waste in rivers. 7,906 annotated images and 27,583 bounding-box annotations across 25 litter categories, from three regions in two countries. COCO detection format.

images	7,906
annotations	27,583
categories	25
regions	3
size	38.1 GB
format	COCO detection (bbox, no masks)

Licence and citation

This dataset is released under the Creative Commons Attribution 4.0 International Licence (CC BY 4.0). The complete legal text is in LICENSE at the root of this repository.

You are free to share (copy and redistribute in any medium or format) and adapt (remix, transform, and build upon the material) for any purpose, including commercially, provided you give appropriate credit, link to the licence, and indicate if changes were made.

Attribution is given by citing the dataset. Under CC BY 4.0 the credit requirement is satisfied by the citation below — include it in any paper, report, model card, or derived dataset that uses this material. No separate permission is needed.

@misc{wu_floatwaste,
  title        = {FloatWaste: a multi-region riverine floating-waste detection benchmark},
  author       = {Wu, Yujiao and PLACEHOLDER-FULL-AUTHOR-LIST},
  howpublished = {PLACEHOLDER-VENUE},
  year         = {PLACEHOLDER-YEAR},
  note         = {Licensed under CC BY 4.0},
  url          = {https://huggingface.co/datasets/OliviaWuYJ/SC-floatwaste}
}

⚠️ The author, howpublished and year fields above are placeholders pending publication and must be replaced with the final reference before the citation is used. The title, licence and URL are final.

This dataset is also available on Zenodo (DOI: 10.5281/zenodo.21934286), which is the citable version of record.

Regions

region	where	images	annotations	per image	size
Hobart	Hobart, Tasmania, Australia	5,413	16,315	3.01	31.7 GB
NSW	Sydney metropolitan councils, NSW, Australia	1,631	1,992	1.22	4.9 GB
Indonesia	Citarum basin, West Java, Indonesia	862	9,276	10.76	1.5 GB
total		7,906	27,583	3.49	38.1 GB

Annotation density differs by almost an order of magnitude across regions: pooled annotation-level metrics are dominated by Indonesia, pooled image-level metrics by Hobart. Indonesia is also the smallest partition at 10.9% of images against Hobart's 68.5%. Report per-region results alongside pooled ones.

Sites

site is published for NSW (council) and Indonesia (location), and is null for Hobart. The Hobart source folders are processing batches with no geographic meaning, so publishing them as sites would invite a spatial reading they cannot support — Hobart imagery was collected uniformly across the Hobart region. source_path retains full provenance for every image regardless.

region	site	images
Indonesia	sapan	283
Indonesia	solokanjeruk	579
NSW	blacktown	111
NSW	canterburybankstown	852
NSW	northernbeaches	149
NSW	parramatta	413
NSW	sopa	106
Hobart	null	5,413

Splits

split	images	annotations	annotation-free images
train	7,110	24,950	49
val	796	2,633	5
total	7,906	27,583	54

Train and val are disjoint and all 25 classes appear in both. The 54 annotation-free images are intentional negatives and are full members of the dataset — see Keep the negatives below.

What is not in the dataset

The Indonesian source folder holds 1,500 images, of which 862 are annotated and released; the other 638 were never annotated and are not part of this dataset — not in any count, split or total on this page. They are raw surplus frames, not negatives. The distinction matters: the 54 negatives are images that were examined and found to contain no litter; the 638 are images nobody has examined. Only the first kind is evidence of absence.

Categories

25 litter classes, ids contiguous 1..25 and identical across both splits.

id	name	anns	id	name	anns
1	Packaging	6,723	14	H_plate/bowl	232
2	Other_packaging	4,960	15	H_utensil	197
3	S_bubblewrap	458	16	DH_lid	779
4	S_label	288	17	D_polystyrene	3,652
5	S_squeeze	475	18	M_beveragecan	651
6	S_straw	174	19	M_foodcan/tin	388
7	PS_string	501	20	M_aerosol	171
8	P_cardboard	1,160	21	R_ball/balloon	394
9	P_foodcontainer	730	22	G_beveragebottle	179
10	PH_cup	1,087	23	F_facemask	51
11	H_packaging	771	24	T_wood/timber	557
12	H_beveragebottle	1,141	25	Other	1,594
13	H_otherbottle	270

The distribution is long-tailed. Packaging (6,723) and Other_packaging (4,960) together account for 42% of all annotations, while F_facemask (51), M_aerosol (171), S_straw (174) fall below 200. This is a property of the litter stream, not a defect, but it should frame any per-class AP discussion.

Directory structure

Images are laid out by region, not by split. The train/val partition lives only in the two annotation JSONs, so re-splitting the dataset never requires moving a file.

floatwaste/
├── README.md
├── LICENSE                  # CC BY 4.0, full legal text
├── annotations/
│   ├── instances_train.json
│   └── instances_val.json
└── images/
    ├── Hobart/          # flat — no site subfolders
    ├── NSW/
    └── Indonesia/

file_name in each JSON is dataset-root-relative, e.g. images/NSW/coxscreek-TIMEL0062_200-717.jpg, so a loader points img_prefix at the dataset root and needs no split-aware path logic. Basenames are unique across the whole dataset.

Metadata schema

Standard COCO, plus three fields on every image record:

field	type	meaning
region	string	Hobart, NSW or Indonesia
site	string or null	council (NSW) or location (Indonesia); null for Hobart
source_path	string	provenance: the image's path in the project's source store

Annotation records are unmodified COCO: bbox is [x, y, width, height] in absolute pixels, area is in square pixels, iscrowd is present, and there are no segmentation masks — this is a detection benchmark.

Loading, and how to filter by region / site

import os
from pycocotools.coco import COCO

ROOT = "."                      # local snapshot of this dataset
coco = COCO(os.path.join(ROOT, "annotations/instances_train.json"))

# every image, with its region and site
for img_id in coco.getImgIds()[:5]:
    info = coco.loadImgs(img_id)[0]
    path = os.path.join(ROOT, info["file_name"])       # -> images/Hobart/....jpg
    anns = coco.loadAnns(coco.getAnnIds(imgIds=img_id))
    print(path, info["region"], info["site"], len(anns))

# filter by region — no path parsing needed
indonesia_ids = [i for i in coco.getImgIds()
                 if coco.loadImgs(i)[0]["region"] == "Indonesia"]

# filter by site (NSW councils / Indonesian locations; Hobart has no site)
parramatta_ids = [i for i in coco.getImgIds()
                  if coco.loadImgs(i)[0]["site"] == "parramatta"]

# annotations for one region, e.g. to score per-region
indonesia_anns = coco.loadAnns(coco.getAnnIds(imgIds=indonesia_ids))

Fetch one region instead of the whole dataset:

from huggingface_hub import snapshot_download

snapshot_download("<org>/<dataset>", repo_type="dataset",
                  allow_patterns=["annotations/*", "images/Indonesia/*"])

Keep the negatives

Some images carry no annotations on purpose — clean water and shoreline retained so a detector learns to produce no detections there. Their emptiness is a label, not a gap, and they are counted in every total on this page. Detection pipelines that drop annotation-free images by default will silently discard them:

# do NOT do this on FloatWaste — it deletes the negative signal
img_ids = [i for i in coco.getImgIds() if coco.getAnnIds(imgIds=i)]

Collection folder naming convention

Collection folders in this project are named:

City_Location_Year_Month_Date               e.g.  Hobart_Rivulet1_2024_October_14
City_Location_BridgeName_Year_Month_Date    e.g.  Bandung_Citarik_Haurpugur_2025_January_20

• Underscores are the only delimiter — no spaces, no hyphens, no commas anywhere in the name.
• BridgeName is used only where a specific bridge is known; otherwise the shorter form applies.
• Date is the capture date of the first image in the folder, not the date the folder was created, copied, or processed.
• Month is the full month name (October, January).

The convention governs NSW, Indonesia, and all future collections. The Hobart material predates it and does not follow it; nothing is parsed from Hobart filenames.

Known limitations

• No capture-date field yet — but many frames are dated in the pixels. No date_captured value is published in this revision, because filenames cannot supply one reliably: NSW filenames encode only camera/sequence/frame, and Hobart files often carry two dates (deployment and frame extraction) that can differ by weeks. The camera-stamped capture date and time is however burned into the imagery itself, in a status strip along the frame edge, for the trail-camera material: all NSW frames (e.g. 10/21/2023 09:03:30AM), the Indonesian frames (e.g. 10/22/2024 07:03:27AM), and the Hobart hobart_scanned subset. The video-derived Hobart frames and the trashnet_reviewedSB images carry no strip. Where both exist they can disagree — one Hobart file named ..._2021_09_21_... is stamped 2021/09/22 — and the burned-in stamp is the capture time. A parsed date field is planned for a future revision. Until then, do not infer capture dates from filenames.
• Image ids are per-split and 0-based in both, so they collide across train and val. Fine for standard COCO loading; any merge must re-key.
• Severe class imbalance and uneven regional composition — see the tables above. Report per-region and per-class results alongside pooled ones.
• A folder name in the provenance is misleading. Some source_path values contain trashnet_reviewedSB. This is self-collected Hobart field data and has nothing to do with the public TrashNet dataset (Thung & Yang) — the name collision is coincidental and carries no provenance or licensing implication.

Provenance and verification

The release was audited and repaired before publication: 977 dead image links repaired, 17 swapped width/height records corrected, and region / site / source_path derived from store paths rather than parsed from filenames. A 30-check verification suite is re-run on every change and must be fully green before release; the annotation and category blocks are byte-identical to the project originals, and 0 bounding boxes are out of bounds or degenerate.

The full audit report and the long-form datasheet are maintained with the project source, not in this repository.

Changelog

v1.0.1 — corrected 1 annotation label found in pre-release stratified review.

One annotation in images/Indonesia/bandung_citarik6_2024_october_12_2_frame_0035.jpg (train, annotation id 24815) was relabelled from R_ball/balloon to Other: the object is an irregular crumpled item, not a ball or balloon. Only category_id changed — the box geometry, area, iscrowd and all ids are untouched, so the annotation total is unchanged at 27,583. Affected class counts: R_ball/balloon 395 → 394, Other 1,593 → 1,594. The annotator confirmed the case is a singleton; adjacent frames were checked and show no repeat. Its attributes.Material was also moved from Rubber to Unknown, since that material had been recorded on the assumption the object was a rubber ball. Label errata are recorded as data in the project source and re-asserted by the verification suite, which fails on any undocumented change to an annotation.

This revision also populates the COCO info and licenses blocks in both annotation files, and repoints every image record's license field to the single CC BY 4.0 entry (id: 1).

v1.0 — initial release: 7,906 images, 27,583 annotations, 25 categories, 3 regions.