Improve Documentation for Loader Output

luxonis_ml/data/loaders/README.md

@@ -6,6 +6,7 @@ The `LuxonisLoader` class provides efficient access to dataset samples with conf
 
 - [LuxonisML Loader](#luxonisml-loader)
   - [Parameters](#parameters)
+  - [Output Structure](#output-structure)
 
 ## Parameters
 
@@ -26,3 +27,49 @@ The `LuxonisLoader` class provides efficient access to dataset samples with conf
 | `keep_categorical_as_strings` | `bool`                                    | `False`            | Whether to keep categorical metadata as strings                                                                                                                                      |
 | `update_mode`                 | `UpdateMode`                              | `UpdateMode.ALL`   | Applicable to remote datasets. The loader internally calls the [`pull_from_cloud`](../datasets/README.md#pulling-from-remote-storage) method to download the dataset from the cloud. |
 | `filter_task_names`           | `Optional[List[str]]`                     | `None`             | If provided, only include annotations for these specified tasks, ignoring any others in the data.                                                                                    |
+
+## Output Structure
+
+`LuxonisLoader` implements `__getitem__`, so each `dataset[i]` returns a **tuple** of `(inputs, labels)`.
+
+```python
+(
+    inputs: np.ndarray | Dict[str, np.ndarray],  # one or more image-like np arrays, where keys are source names
+    labels: Labels                               # task-specific labels
+)
+```
+
+### Accessing Labels
+
+Labels can be accessed using specific keys in the `labels` dictionary. These keys follow the format:
+
+```python
+f"{task_name}/{task_type}"
+```
+
+If the dataset was created without specifying a `task_name`, the default keys will be:
+
+```
+/boundingbox, /classification, /segmentation, /instance_segmentation, /keypoints, /metadata/<key>
+```
+
+In the case of metadata, the `/metadata/<key>` format uses a field name that was provided when creating the dataset in place of `<key>` (e.g., `/metadata/id`, `/metadata/camera_angle`, etc.).
+
+If a dataset was created using multiple `task_name`s—which is especially helpful for more structured or complex datasets—you might define one task for **segmentation** and another for **keypoint detection**. In that case, you would access the labels using keys like:
+
+```
+segmentation_task/segmentation, pose_task/keypoints
+```
+
+This naming convention makes it easier to manage, access, and process annotations in multi-task datasets.
+
+### Labels Structure
+
+| Task Type                 | Example Shape                    | Description                                                                                                              |
+| ------------------------- | -------------------------------- | ------------------------------------------------------------------------------------------------------------------------ |
+| **classification**        | `(N_classes,)`                   | One-hot encoded classification vector.                                                                                   |
+| **segmentation**          | `(C, H, W)`                      | One-hot encoded per-pixel class mask in CHW format, where `C` is the number of classes.                                  |
+| **boundingbox**           | `(N_instances, 5)`               | Each row represents an object as `[cls, x_min, y_min, w, h]`.                                                            |
+| **instance_segmentation** | `(N_instances, H, W)`            | Binary mask for each instance.                                                                                           |
+| **keypoints**             | `(N_instances, n_keypoints * 3)` | Keypoints in `x, y, v` format (visibility flag), repeated per keypoint.                                                  |
+| **metadata**              | —                                | Follows the same structure as yielded. Values are accessible via keys such as `"metadata/key1"`, `"metadata/key2"`, etc. |