4w4kt/waste-dataset-image

♻️ Waste Classification Image Dataset

Dataset Summary

This dataset contains standardized, multi-class waste images categorized across 9 target recycling categories. It was compiled, curated, and manually sanitized to eliminate noise, corrupted files, and class overlap in order to train high-performance convolutional neural networks and transfer learning backbones (such as EfficientNet-B5).

The dataset contains a total of 28,840 labeled images, split into dedicated training and testing subsets, with an optimized Parquet download footprint of ~6.4 GB.

---

👥 Authorship & Curation

This dataset was assembled, verified, and manually cleaned through collaborative work by:

• Alejandra Rodríguez Silva (@alejandra-rs / GitHub)
• José Marcial Galván Franco (@4w4kt / GitHub)

Manual Cleaning & Preprocessing

• Manual Sanitization: Images sourced across multiple raw collections were manually screened to eliminate duplicates, noisy/ambiguous items, and mislabeled classes.
• Resolution & Normalization: Input resolution normalization and adaptive resizing routines (including 456x456 and 528x528 dimensions matching EfficientNet compound scaling).
• Class Balancing: Class-weighted cross-entropy loss functions were designed alongside this distribution to mitigate sample imbalances across bins (such as battery and light bulb waste vs. clothes).

---

🌐 Sourced Datasets (Origins)

The base imagery originates from a synthesis of 7 open-source collections on Kaggle:

1. Custom Waste Classification Dataset - Wasif Mahmood
2. Garbage Dataset - Suman Kunwar
3. Waste Pictures - 且听风吟
4. Waste Classification Dataset - Kaan Çerkez
5. Diverse Tools Image Dataset for Machine Learning - Oort Datahub
6. Garbage Dataset from Various Sources - Butleriii
7. Object Detection: Batteries, Dice, and Toy Cars - Márk Antal Csizmadia

---

📊 Dataset Structure

Partitions & Size

Split	Number of Images	Download Size
train	23,264	~5.8 GB
test	5,576	~0.6 GB
Total	28,840	~6.4 GB

Classes & Indices

ID	Class Label	Category Focus
0	E-waste	Small appliances, circuit boards, components
1	battery waste	Alkaline, lithium, cylindrical, and car batteries
2	clothes	Textiles, garments, fabrics
3	glass waste	Bottles, jars, broken glass fragments
4	light bulbs	Incandescent, LED, fluorescent tubes
5	metal waste	Aluminum cans, scrap metal, tins
6	organic waste	Food leftovers, peels, biodegradable waste
7	paper waste	Cardboard boxes, paper sheets, packaging
8	plastic waste	PET bottles, containers, bags

---

💻 Usage

Load the data directly through the Hugging Face datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("4w4kt/waste-dataset-image")

# Inspect a train sample
sample = dataset["train"][0]
image = sample["image"]
label_id = sample["label"]
label_name = dataset["train"].features["label"].int2str(label_id)

print(f"Sample Class: {label_name} (ID: {label_id})")

---

🔗 Related Models & Implementations

• Pretrained Weights: 4w4kt/waste-classifier-b5
• Training Code & Architecture Study: GitHub Repository