CLICK Trash Dataset

Published: 1 November 2026| Version 1 | DOI: 10.17632/htcvvz6fm2.1
Contributor:

Description

CLICK TRASH DATASET =================== Real litter, real ground, real clutter. 1,126 outdoor photos with 2,213 polygon-labelled pieces of trash, in YOLO and COCO formats, with two label spaces: 1 class (trash) and 13 material classes. ANNOTATION Masks were first proposed by Segment Anything Model 3 (SAM 3 by Meta AI). Every image was then checked and corrected by hand in Label Studio. No annotation is unreviewed model output. CONTENTS - Scenes: soil, leaf litter, grass, gravel, paths; phone photos in natural light, mostly 3072 x 3072 px - Split : train 901 / val 113 / test 112 images and corresponding 1,802 / 221 / 190 objects - About 2 objects per image (max 29) - A total of 4 background images with no litter - Supports object detection and instance segmentation LABEL SPACES (same images, split and polygons) - single class: trash - multi class (13): Bottle Cap, Cardboard, Crumbled Paper, Match Box, Metal Can, Others, Paper Cup, Paper Pack, Plastic Bag, Plastic Bottle, Plastic Packet, Silver Foil, Slippers CLASS BALANCE Long-tailed: Plastic Packet is 1,112 of 2,213 objects; Match Box, Slippers, Cardboard and Bottle Cap have under 15 each. Bottle Cap and Slippers are absent from test; Match Box from val and test. Consider class-weighted or focal loss, oversampling, or merging rare classes into Others, and report per-class metrics. FOLDERS images/{train,val,test} shared photos, named trash (N).jpg single class/, multi class/ each with labels/ (YOLO), annotations/ (COCO instances_{split}.json), classes.txt, data.yaml FORMATS YOLO: class cx cy w h, normalised 0-1 (tight box of the COCO polygon). COCO: polygon segmentation, bbox [x, y, w, h] in pixels. YOLO ids start at 0, COCO ids at 1. TRAINING NOTE Ultralytics needs images/ and labels/ side by side, so build a working folder of per-file symlinks (images from images/<split>, labels from <label space>/labels/<split>) and point data.yaml at it. QUALITY CHECKS Every image has a label file and COCO entry; boxes lie inside images with no duplicates; YOLO matches COCO within 0.1%; image sizes match COCO; both label spaces share images, split and geometry. CITATION Baishya, Manjit (2026). CLICK Trash Dataset, v1.0. Mendeley Data. DOI: 10.17632/473x7yf3vy.1 https://doi.org/10.17632/473x7yf3vy.1 GitHub: to be added. License: to be added. Report label errors with the image name, e.g. trash (362).jpg. ACKNOWLEDGEMENTS SAM 3 (Meta AI); Label Studio (HumanSignal). CLICK: you can't clean up what you can't see.

Files

Categories

Image Segmentation, Object Detection, YOLOv5

Licence