larmkaixian/malaysia-trash-annotation

Notice:

The foundational manuscript detailing the methodology, taxonomy (Build 20260418), and benchmark performance of the MTA dataset is currently Under Review. If you are utilizing this dataset for research, please bookmark this repository. The official BibTeX citation will be published here upon the paper's acceptance.

malaysian-trash-annotation > Build-20260418-x1

-x1 refer to dataset without any augmentation applied to.

Check out the Roboflow Universe

License: CC BY 4.0

Malaysian Trash Annotation (MTA)

Check out the Github repository: Main MTA Repository

--- Acknowledgement ---

We thank all contributors for their support in data collection and annotation: Daniel Looi Jun Jie, Darren Lau, Ice019, Ken, Kevin Tan, Kevin Tan Feng Sheng, Lim Li Hong, Nazzy, Ng Lee Pink, Tey KY, Trash, cyarah >~<

--- Introduction ---

Malaysian Trash Annotation or MTA in short is a localised instance segmentation dataset designed to address the current issue of waste segregation in Southeast Asia context. Compare with other openly-sourced dataset, MTA is heavily designed around real-world situation where it provides polygon masking to capture the exact geometry of the trash.

--- Publication ---

If you use our dataset or technical report, please cite us using:

@dataset{mta2026, author = {Larm, K. X. and Lawi, A.}, title = {Malaysian Trash Annotation (MTA)}, year = {2026}, doi = {10.5281/zenodo.17862878}, url = {https://huggingface.co/datasets/larmkaixian/malaysia-trash-annotation}, note = {Dataset; formal manuscript in preparation} }

--- Goal ---

To develop a high-accuracy waste segmentation model that can precisely localize objects with mainly account for real-world Malaysian environments.

--- Dataset's Version: Build 20260418 ---

• Classes: 14 (Material-based separation)
• Environment: Supermarkets (clean and ideal condition), Roadside (weathered), Food Courts (occluded/deformed).
• Hardware: Heterogeneous capture using 5 different devices (Oppo, Honor, Samsung, iPhone, Realme).
• Capture Protocol: Aspect Ratio locked at 4:3, Multi-angle (Top, Side, 45-degree) with no AI enhancements and no HDR enabled.

Note 1:

We overhaul the dataset's structure to now only support for material-based detection, should you split into 3 different segment in order to reach maximum accuracy of the purpose of detecting the recyclability of the trash, which are material-based, contamination-status, and object-based approach. For more information, please refer to Singapore National Environment Agency guideline here.

Note 2:

Partial reference are taken from Singapore National Environment Agency guideline and from Solid Waste Management and Public Cleansing Corperation (SWCorp) Malaysia

--- Annotation Strategy ---

The dataset is currently structured for material-based detection (Phase 1), as below:

Future extensions include:

Object-level classification Contamination detection

This modular approach aligns with practical waste classification systems.

--- Taxonomy of dataset ---

m-* (Main Category)

Taxonomy	Description
m-composite	refer to such trash has composite material with it
m-glass	refer to such trash is mainly made out of glass
m-metal	refer to such trash is mainly made out of metal
m-paper	refer to such trash is mainly made out of paper
m-plastic-film	refer to such trash is mainly made out of soft plastic
m-plastic-rigid	refer to such trash is mainly made out of hard plastic

s-* (Special Category)

Taxonomy	Description
s-e-waste	refer to such trash is e-waste
s-hazardous	refer to such trash is hazardous or dangerous to environment and human
s-ikat-tepi	refer to such trash is ikat tepi
s-litter	refer to such trash is common trash and can't be recycle whatsoever
s-organic	refer to such trash is organic waste
s-other	refer to such trash can't be categorized or too little to become its own class
s-textile	refer to such trash is made out of textile
s-cigarette-butt	refer to such trash is cigarette butt

--- Devices ---

• Oppo Reno 14 MY Version (OppoReno14)
• Samsung Note 10 MY Version (SamsungNote10)
• Iphone 11 MY Version (Iphone11)
• Honor 200 MY Version (Honor200)
• Realme 5s MY Version (Realme5s)

--- References ---

Singapore National Environment Agency (NEA): [https://www.nea.gov.sg/docs/default-source/our-services/waste-management/list-of-items-that-are-recyclable-and-not.pdf]

SWCorp Malaysia: [https://www.swcorp.gov.my/asingkan/]

--- Legal & Disclaimer ---

This dataset contains real-world images of public environments. Any trademarks, logos, or brand names visible on objects (e.g., packaging, bottles) are the property of their respective owners.

Such elements are captured incidentally as part of natural scenes and are included solely for research, educational, and computer vision development purposes. The authors do not claim ownership of any third-party trademarks, nor does their inclusion imply any affiliation with or endorsement by the respective rights holders.

The dataset is provided “as is” without warranties, and the authors are not responsible for any misuse of the data.

Notes:

• Reflections: Contains samples on glass/steel tables; polygons are tightly cropped to the object, excluding ghost reflections.
• Deformation: Includes crushed cans and flattened bottles to simulate end-of-life object states.
• Localization: Features "Ikat Tepi" and "Keropok" wrappers specific to the Malaysian region.
• Build-20260418 may mark as final version for MTAv1 as MTAv2 introduce fundemental different architecture which MTAv2 will utilizing multi-attribute system.