---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: mask
    dtype: image
  - name: class_id
    dtype:
      class_label:
        names:
          '0': verde_plastico
          '1': azul
          '2': negro_plastico
          '3': negro_carton
          '4': roja_plastico
          '5': carton
          '6': verde_carton
          '7': roja_eldulze
          '8': verde_plastico_oscuro
          '9': verde_cogollo
          '10': ilfres
  - name: bbox
    sequence: float64
  - name: mask_rgb_color
    sequence: int64
  splits:
  - name: loads
    num_bytes: 303854470
    num_examples: 949
  download_size: 292816441
  dataset_size: 303854470
configs:
- config_name: default
  data_files:
  - split: loads
    path: data/loads-*
license: mit
task_categories:
- object-detection
- image-segmentation
size_categories:
- n<1K
tags:
- industry
---

The **IndustrialLateralLoads** dataset is designed for object detection and instance segmentation tasks in industrial environments. It contains images of palletized loads with their corresponding annotations. The dataset is available in two formats:
- **Hugging Face dataset (Parquet):** Ready-to-use format with images, masks, and metadata.
- **Raw files:** Original folders accessible in the repository files.

### Hugging Face dataset features:
When loaded using the `datasets` library, each sample contains the following fields:

- **`image`**: A `PIL.Image` object containing the original RGB image.
- **`mask`**: A `PIL.Image` object containing the binary segmentation mask.
- **`class_id`**: The classification label of the load (integer mapped to class name).
- **`bbox`**: The bounding box coordinates in format `[x_min, y_min, width, height]`.
- **`mask_rgb_color`**: An `[R, G, B]` list specifying the RGB triplet associated with the class in the binary mask.

### Raw data organization (repository files):
If you prefer to download the raw files, the repository includes the following folders:

- **`images`**: Contains the source images.
- **`masks`**: Contains the binary PNG segmentation masks.
- **`bboxes`**: Contains `.txt` files with bounding box annotations.

**Remark 1.** Each row in a `.txt` file (inside the `bboxes` folder) follows this format:

`<class_name> <instance_id> <x_min> <y_min> <width> <height>`,

where the `<instance_id>` field is not relevant in the currently published dataset.

**Remark 2.** The segmentation masks are provided as **binary RGB images**. They contain only **two** distinct pixel values:
- **Background:** Black `(0, 0, 0)`.
- **Foreground:** The specific **RGB Color** associated with the class ID listed below.
| Class ID | Class Name              | Foreground RGB Color |
| :------- | :---------------------- | :------------------- |
| **0** | **verde_plastico** | `(51, 221, 255)`  |
| **1** | **azul** | `(255, 204, 51)`  |
| **2** | **negro_plastico** | `(52, 209, 183)`  |
| **3** | **negro_carton** | `(255, 96, 55)`   |
| **4** | **roja_plastico** | `(36, 179, 83)`   |
| **5** | **carton** | `(255, 125, 187)` |
| **6** | **verde_carton** | `(221, 255, 51)`  |
| **7** | **roja_eldulze** | `(170, 240, 209)` |
| **8** | **verde_plastico_oscuro**| `(255, 10, 124)` |
| **9** | **verde_cogollo** | `(184, 61, 245)`  |
| **10** | **ilfres** | `(116, 113, 206)` |
### Scientific validation & ongoing research:
This dataset serves as a benchmark for validating computer vision approaches in industrial logistics. It remains an active resource for ongoing research, with current efforts shifting towards **instance segmentation** methodologies in industrial environments.

##### **Associated publications:**
The dataset has been employed to validate the approaches presented in the following scientific studies:

* **[Fog Computing-Driven Logistics: Leveraging Few-Shot Learning and Foundational Computer Vision Models](https://link.springer.com/article/10.1007/s10586-025-05662-w)**

* **[Enhancing Logistics with Computer Vision and Fog Computing-Driven Auto-ID Technologies](https://ieeexplore.ieee.org/document/10710204)**

* **[From Few-Shot to Zero-Shot Pallet Load Recognition: A Deployed Embedding-Based Vision System for Industrial Logistics](https://ieeexplore.ieee.org/document/11492154)**
    - *Accepted at WACV 2026 (Winter Conference on Applications of Computer Vision)*

* **Training-Free One-Shot Industrial Load Segmentation for Real-Time Logistics Operations**
    - *Under Review*
 
---

### Install Hugging Face datasets package:
```sh
pip install datasets
```

### Usage example:
```python
from datasets import load_dataset

# Load the dataset (images and masks are loaded as PIL objects)
dataset = load_dataset("jjldo21/IndustrialLateralLoads")

# Example: Accessing the first sample
sample = dataset['loads'][0]
print(sample['class_id'])       # Class label (e.g., 9 corresponding to 'verde_cogollo')
print(sample['mask_rgb_color']) # Color of the mask (e.g., [184, 61, 245])
sample['image'].show()          # Display image
sample['mask'].show()           # Display mask
```