Update README.md

README.md

@@ -56,4 +56,54 @@ configs:
     path: data/train-*
   - split: val
     path: data/val-*
+task_categories:
+- depth-estimation
+- image-segmentation
+- image-feature-extraction
+size_categories:
+- 1K<n<10K
 ---
+
+
+This is the NYUv2 dataset for scene understanding tasks.
+I downloaded the original data from the [Tsinghua Cloud](https://cloud.tsinghua.edu.cn/f/6d0a89f4ca1347d8af5f/?dl=1) and transformed it into Huggingface Dataset.
+Credit to [ForkMerge: Mitigating Negative Transfer in Auxiliary-Task Learning](http://arxiv.org/abs/2301.12618).
+
+## Dataset Information
+
+This data contains two splits: 'train' and 'val' (used as test dataset).
+Each sample in the dataset has 5 items: 'image', 'segmentation', 'depth', 'normal', and 'noise'.
+The noise is generated using `torch.rand()`.
+
+
+## Usage
+
+```python
+dataset = load_dataset('tanganke/nyuv2')
+dataset = dataset.with_format('torch') # this will convert the items into `torch.Tensor` objects
+```
+
+this will return a `DatasetDict`:
+
+```python
+DatasetDict({
+    train: Dataset({
+        features: ['image', 'segmentation', 'depth', 'normal', 'noise'],
+        num_rows: 795
+    })
+    val: Dataset({
+        features: ['image', 'segmentation', 'depth', 'normal', 'noise'],
+        num_rows: 654
+    })
+})
+```
+
+The features:
+
+```python
+{'image': Array3D(shape=(3, 288, 384), dtype='float32', id=None),
+ 'segmentation': Array2D(shape=(288, 384), dtype='int64', id=None),
+ 'depth': Array3D(shape=(1, 288, 384), dtype='float32', id=None),
+ 'normal': Array3D(shape=(3, 288, 384), dtype='float32', id=None),
+ 'noise': Array3D(shape=(1, 288, 384), dtype='float32', id=None)}
+```