Imagenet-s dataset for large-scale semantic segmentation (#2480)

## Motivation

Based on the ImageNet dataset, we propose the ImageNet-S dataset has 1.2 million training images and 50k high-quality semantic segmentation annotations to support unsupervised/semi-supervised semantic segmentation on the ImageNet dataset.

paper:
Large-scale Unsupervised Semantic Segmentation (TPAMI 2022)
[Paper link](https://arxiv.org/abs/2106.03149)

## Modification

1. Support imagenet-s dataset and its' configuration
2. Add the dataset preparation in the documentation

README.md

@@ -188,6 +188,7 @@ Supported datasets:
 - [x] [Vaihingen](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md#isprs-vaihingen)
 - [x] [iSAID](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md#isaid)
 - [x] [High quality synthetic face occlusion](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md#delving-into-high-quality-synthetic-face-occlusion-segmentation-datasets)
+- [x] [ImageNetS](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md#imagenets)
 
 ## FAQ
 

configs/_base_/datasets/imagenets.py

@@ -0,0 +1,61 @@
+# dataset settings
+dataset_type = 'ImageNetSDataset'
+subset = 919
+data_root = 'data/ImageNetS/ImageNetS919'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+crop_size = (224, 224)
+train_pipeline = [
+    dict(type='LoadImageNetSImageFromFile', downsample_large_image=True),
+    dict(type='LoadImageNetSAnnotations', reduce_zero_label=False),
+    dict(type='Resize', img_scale=(1024, 256), ratio_range=(0.5, 2.0)),
+    dict(
+        type='RandomCrop',
+        crop_size=crop_size,
+        cat_max_ratio=0.75,
+        ignore_index=1000),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=1000),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageNetSImageFromFile', downsample_large_image=True),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1024, 256),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type=dataset_type,
+        subset=subset,
+        data_root=data_root,
+        img_dir='train-semi',
+        ann_dir='train-semi-segmentation',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        subset=subset,
+        data_root=data_root,
+        img_dir='validation',
+        ann_dir='validation-segmentation',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        subset=subset,
+        data_root=data_root,
+        img_dir='validation',
+        ann_dir='validation-segmentation',
+        pipeline=test_pipeline))

docs/en/dataset_prepare.md

@@ -155,6 +155,25 @@ mmsegmentation
 │   │   │   ├── img
 │   │   │   ├── mask
 │   │   │   ├── split
+│   ├── ImageNetS
+│   │   ├── ImageNetS919
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+│   │   ├── ImageNetS300
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+│   │   ├── ImageNetS50
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
 ```
 
 ### Cityscapes
@@ -580,3 +599,31 @@ OCCLUDER_DATASET.IMG_DIR "path/to/jw93/mmsegmentation/data_materials/DTD/images"
 ```python
 
 ```
+
+### ImageNetS
+
+The ImageNet-S dataset is for [Large-scale unsupervised/semi-supervised semantic segmentation](https://arxiv.org/abs/2106.03149).
+
+The images and annotations are available on [ImageNet-S](https://github.com/LUSSeg/ImageNet-S#imagenet-s-dataset-preparation).
+
+```
+│   ├── ImageNetS
+│   │   ├── ImageNetS919
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+│   │   ├── ImageNetS300
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+│   │   ├── ImageNetS50
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+```

docs/zh_cn/dataset_prepare.md

@@ -119,6 +119,25 @@ mmsegmentation
 │   │   ├── ann_dir
 │   │   │   ├── train
 │   │   │   ├── val
+│   ├── ImageNetS
+│   │   ├── ImageNetS919
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+│   │   ├── ImageNetS300
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+│   │   ├── ImageNetS50
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
 ```
 
 ### Cityscapes
@@ -317,3 +336,31 @@ python tools/convert_datasets/isaid.py /path/to/iSAID
 ```
 
 使用我们默认的配置 (`patch_width`=896, `patch_height`=896,　`overlap_area`=384)， 将生成 33978 张图片的训练集和 11644 张图片的验证集。
+
+### ImageNetS
+
+ImageNet-S是用于[大规模无监督/半监督语义分割](https://arxiv.org/abs/2106.03149)任务的数据集。
+
+ImageNet-S数据集可在[ImageNet-S](https://github.com/LUSSeg/ImageNet-S#imagenet-s-dataset-preparation)获取。
+
+```
+│   ├── ImageNetS
+│   │   ├── ImageNetS919
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+│   │   ├── ImageNetS300
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+│   │   ├── ImageNetS50
+│   │   │   ├── train-semi
+│   │   │   ├── train-semi-segmentation
+│   │   │   ├── validation
+│   │   │   ├── validation-segmentation
+│   │   │   ├── test
+```

mmseg/datasets/__init__.py

@@ -11,6 +11,8 @@ from .dataset_wrappers import (ConcatDataset, MultiImageMixDataset,
 from .drive import DRIVEDataset
 from .face import FaceOccludedDataset
 from .hrf import HRFDataset
+from .imagenets import (ImageNetSDataset, LoadImageNetSAnnotations,
+                        LoadImageNetSImageFromFile)
 from .isaid import iSAIDDataset
 from .isprs import ISPRSDataset
 from .loveda import LoveDADataset
@@ -27,5 +29,7 @@ __all__ = [
     'PascalContextDataset59', 'ChaseDB1Dataset', 'DRIVEDataset', 'HRFDataset',
     'STAREDataset', 'DarkZurichDataset', 'NightDrivingDataset',
     'COCOStuffDataset', 'LoveDADataset', 'MultiImageMixDataset',
-    'iSAIDDataset', 'ISPRSDataset', 'PotsdamDataset', 'FaceOccludedDataset'
+    'iSAIDDataset', 'ISPRSDataset', 'PotsdamDataset', 'FaceOccludedDataset',
+    'ImageNetSDataset', 'LoadImageNetSAnnotations',
+    'LoadImageNetSImageFromFile'
 ]