README.md

---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': cardboard
          '1': glass
          '2': metal
          '3': paper
          '4': plastic
          '5': trash
  - name: label_id
    dtype: int64
  - name: label_name
    dtype: string
  - name: source_dataset
    dtype: string
  - name: original_label
    dtype: string
  - name: image_context
    dtype: string
  splits:
  - name: train
    num_bytes: 1450173709
    num_examples: 20894
  - name: validation
    num_bytes: 282857438
    num_examples: 4477
  - name: test
    num_bytes: 279046679
    num_examples: 4478
  download_size: 2068691097
  dataset_size: 2012077826
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: mit
task_categories:
- image-classification
language:
- en
tags:
- waste-classification
- trash-classification
- image-classification
size_categories:
- 1K<n<10K
---
# Dataset Card for Waste Classification Harmonized Dataset

## Dataset Details

### Dataset Description

This dataset is a harmonized image classification dataset for waste classification. It combines multiple waste-related image datasets into a unified label structure with six target classes:

* `cardboard`
* `glass`
* `metal`
* `paper`
* `plastic`
* `trash`

The dataset is designed for training and evaluating computer vision models for waste classification, especially in the context of recycling and real-world trash recognition. The original datasets used different label names and different levels of class granularity, so their labels were mapped into one common classification scheme.

The final dataset contains **29,849 images**.

* **License:** MIT
  
### Dataset Sources

This project combines and harmonizes the following public waste-classification datasets:

#### TrashNet

* **Source:** Kaggle — TrashNet by Feyza Ozkefe
* **Link:** https://www.kaggle.com/datasets/feyzazkefe/trashnet
* **Notes:** Contains six TrashNet-style classes: `cardboard`, `glass`, `metal`, `paper`, `plastic`, and `trash`.
* **Used for:** Base label structure and additional image data.
* **License:** MIT

#### Household Garbage Dataset

* **Source:** Kaggle — Recyclable and Household Waste Classification
* **Link:** https://www.kaggle.com/datasets/alistairking/recyclable-and-household-waste-classification/data
* **Notes:** Contains household garbage classes with `default` and `real_world` image contexts.
* **Used for:** Additional training data and real-world robustness evaluation.
* **License:** MIT

#### The Garbage Dataset / Garbage Classification V2

* **Source:** Kaggle — Garbage Classification V2
* **Link:** https://www.kaggle.com/datasets/sumn2u/garbage-classification-v2
* **Notes:** Only the `original` folder was used. The `standardized_256` and `standardized_384` folders were ignored to avoid duplicate or near-duplicate images.
* **Used for:** Additional training data after label harmonization.
* **License:** MIT

#### TriCascade WasteImage

* **Source:** Kaggle — TriCascade WasteImage Dataset
* **Link:** https://www.kaggle.com/datasets/salam035/waste-datasets
* **Used for:** Additional training data after label harmonization.
* **License:** MIT
## Uses

### Direct Use

This dataset can be used for:

* image classification
* waste classification
* recycling-related computer vision experiments
* transfer learning
* fine-tuning pretrained vision models
* comparing image classification architectures
* evaluating model robustness across different image sources

Suitable model architectures include, but are not limited to:

* ResNet
* ConvNeXt
* DeiT
* EfficientNet
* MLP-Mixer
* Vision Transformer variants

### Out-of-Scope Use

This dataset is not intended for direct deployment in production-level recycling systems without further validation.

The dataset should not be used as the only evaluation source for real-world waste sorting systems, because real-world environments may include:

* multiple objects in one image
* damaged or dirty waste objects
* unusual lighting conditions
* occlusion
* mixed materials
* unseen waste categories
* local recycling rules that differ by region

The dataset is also not suitable for fine-grained waste classification such as distinguishing between PET, HDPE, organic waste, batteries, electronic waste, or hazardous materials.

## Dataset Structure

Each sample contains an image, a harmonized label, and metadata about the original source.

Example sample:

```python
{
    "image": "<PIL.Image.Image>",
    "label": 4,
    "label_id": 4,
    "label_name": "plastic",
    "source_dataset": "household_garbage",
    "original_label": "plastic_soda_bottles",
    "image_context": "real_world"
}
```

### Fields

| Field            | Description                                                                                          |
| ---------------- | ---------------------------------------------------------------------------------------------------- |
| `image`          | Image file loaded as a Hugging Face image feature                                                    |
| `label`          | Integer class label used for training                                                                |
| `label_id`       | Integer ID of the harmonized label                                                                   |
| `label_name`     | Name of the harmonized class                                                                         |
| `source_dataset` | Name of the dataset from which the image originated                                                  |
| `original_label` | Original label before harmonization                                                                  |
| `image_context`  | Context information about the image, for example `real_world`, `default`, or source-specific context |

### Label Mapping

| Label ID | Label Name |
| -------: | ---------- |
|        0 | cardboard  |
|        1 | glass      |
|        2 | metal      |
|        3 | paper      |
|        4 | plastic    |
|        5 | trash      |

### Label Distribution

The final harmonized dataset contains **29,849 images** across six classes.

| Label     | Count |
| --------- | ----: |
| cardboard | 4,849 |
| glass     | 5,000 |
| metal     | 5,000 |
| paper     | 5,000 |
| plastic   | 5,000 |
| trash     | 5,000 |

Most classes were limited to **5,000 images per label** to reduce class imbalance. The `cardboard` class contains **4,849 images**, because fewer images were available after harmonization and filtering.

### Data Splits

| Split      | Number of Images |
| ---------- | ---------------: |
| train      |           20,894 |
| validation |            4,477 |
| test       |            4,478 |

Total: **29,849 images**

## Dataset Creation

### Curation Rationale

Waste classification datasets often use different class names, different label granularities, and different image conditions. For example, one dataset may use a broad label such as `plastic`, while another may contain more specific labels such as `plastic_soda_bottles`, `plastic_food_containers`, or `plastic_cup_lids`.

The goal of this dataset is to create a unified and practical waste-classification dataset with a common label structure. This makes it easier to train and compare different image classification models under the same classification task.

### Source Data

The source data consists of waste images from multiple public Kaggle datasets. These datasets contain images of recyclable and non-recyclable waste objects, including cardboard, glass, metal, paper, plastic, and general trash.

The original datasets use different label names, image contexts, and levels of class granularity. Therefore, their labels were mapped into one shared six-class structure:

* `cardboard`
* `glass`
* `metal`
* `paper`
* `plastic`
* `trash`

The original dataset name and original label are preserved in the metadata fields `source_dataset` and `original_label`.

### Annotations

The dataset does not introduce completely new manual annotations for every image. Instead, it harmonizes existing source labels into a shared six-class label structure.

#### Annotation Process

The annotation process consisted mainly of label harmonization. Source-specific labels were mapped to one of the following final classes:

* `cardboard`
* `glass`
* `metal`
* `paper`
* `plastic`
* `trash`

For example:

* `plastic_soda_bottles` → `plastic`
* `plastic_food_containers` → `plastic`
* `plastic_cup_lids` → `plastic`

The original labels are preserved in the dataset metadata.

#### Who are the annotators?

The original annotations were provided by the creators of the source datasets. The harmonized label mapping was created during the dataset preparation process.

#### Personal and Sensitive Information

This dataset contains images of waste objects. It is not intended to contain personal, sensitive, or private information.

However, because the dataset contains real-world images from multiple sources, there is a possibility that some images may include background objects or environmental details. Users should inspect the dataset before using it in privacy-sensitive applications.

## Bias, Risks, and Limitations

This dataset has several limitations:

1. **Source bias**
   The dataset combines images from different sources. Each source may have different image styles, backgrounds, camera qualities, object positions, and lighting conditions.

2. **Background bias**
   Some images may contain clean or artificial backgrounds, while others may contain more realistic environments. A model trained on this dataset may learn background patterns instead of only learning the waste object.

3. **Label noise**
   Some images may contain multiple objects, unclear objects, damaged waste items, or objects made of mixed materials. In such cases, a single label may not fully describe the image.

4. **Limited taxonomy**
   The dataset only contains six broad classes. It does not distinguish between more specific waste categories such as organic waste, batteries, electronics, hazardous materials, or different plastic types.

5. **Near-duplicate images**
   Since the dataset was created from multiple sources, there may be duplicate or near-duplicate images. Users should perform duplicate detection if they need a strict evaluation setup.

6. **Real-world generalization**
   Good performance on this dataset does not guarantee good performance in real recycling facilities or outdoor waste-detection scenarios.

### Recommendations

Users should evaluate models with more than accuracy alone. Recommended metrics include:

* macro F1-score
* per-class precision
* per-class recall
* confusion matrix
* validation loss
* test loss

Macro F1-score is especially recommended because each waste class should be treated equally.

For stronger evaluation, users should consider source-aware testing. For example, they can train on some source datasets and test on a different source dataset to measure domain generalization.
**APA:**

Lavassani, A. (2026). *Waste Classification Harmonized Dataset*. Hugging Face.

## Glossary

**Harmonization**
The process of mapping labels from different datasets into one shared label structure.

**Source dataset**
The original dataset from which an image was taken.

**Original label**
The label assigned to the image in its original dataset before harmonization.

**Image context**
Metadata describing the visual context or source-specific image condition, such as real-world or default images.

**Macro F1-score**
An evaluation metric that calculates the F1-score for each class separately and then averages them equally. This is useful when all classes are important.

## More Information

This dataset was created as part of a machine learning project for waste image classification. The project focuses on dataset harmonization, model comparison, transfer learning, and fine-tuning pretrained computer vision models.