【Hackathon + No.150】PaddleRS集成PaddleDetection的旋转框检测能力

docs/apis/data_cn.md

@@ -57,7 +57,7 @@
 |-------|----|--------|-----|
 |`data_dir`|`str`|数据集存放目录。||
 |`image_dir`|`str`|输入图像存放目录。||
-|`ann_path`|`str`|[COCO格式](https://cocodataset.org/#home)标注文件路径。||
+|`anno_path`|`str`|[COCO格式](https://cocodataset.org/#home)标注文件路径。||
 |`transforms`|`paddlers.transforms.Compose`|对输入数据应用的数据变换算子。||
 |`label_list`|`str` \| `None`|label list文件。label list是一个文本文件，其中每一行包含一个类别的名称。|`None`|
 |`num_workers`|`int` \| `str`|加载数据时使用的辅助进程数。若设置为`'auto'`，则按照如下规则确定使用进程数：当CPU核心数大于16时，使用8个数据读取辅助进程；否则，使用CPU核心数一半数量的辅助进程。|`'auto'`|

docs/apis/data_en.md

@@ -57,7 +57,7 @@ The initialization parameter list is as follows:
 |-------|----|--------|-----|
 |`data_dir`|`str`|Directory that stores the dataset.||
 |`image_dir`|`str`|Directory of input images.||
-|`ann_path`|`str`|[COCO Format](https://cocodataset.org/#home)label file path.||
+|`anno_path`|`str`|[COCO Format](https://cocodataset.org/#home)label file path.||
 |`transforms`|`paddlers.transforms.Compose`|Data transformation operators applied to input data.||
 |`label_list`|`str` \| `None`|Label list path. Label list is a text file, in which each line contains the name of class.|`None`|
 |`num_workers`|`int` \| `str`|Number of auxiliary processes used when loading data. If it is set to `'auto'`, use the following rules to determine the number of processes to use: When the number of CPU cores is greater than 16, 8 data read auxiliary processes are used; otherwise, the number of auxiliary processes is set to half the counts of CPU cores.|`'auto'`|

docs/apis/train_cn.md

@@ -156,6 +156,8 @@ def train(self,
           train_dataset,
           train_batch_size=64,
           eval_dataset=None,
+          train_batch_transforms=None,
+          eval_batch_transforms=None,
           optimizer=None,
           save_interval_epochs=1,
           log_interval_steps=10,
@@ -186,6 +188,8 @@ def train(self,
 |`train_dataset`|`paddlers.datasets.COCODetDataset` \| `paddlers.datasets.VOCDetDataset` |训练数据集。||
 |`train_batch_size`|`int`|训练时使用的batch size。|`64`|
 |`eval_dataset`|`paddlers.datasets.COCODetDataset` \| `paddlers.datasets.VOCDetDataset` \| `None`|验证数据集。|`None`|
+|`train_batch_transforms`|`paddlers.transforms.BatchCompose` \| `None`|训练集上的批数据变换算子。|`None`|
+|`eval_batch_transforms`|`paddlers.transforms.BatchCompose` \| `None`|验证集上的批数据变换算子。|`None`|
 |`optimizer`|`paddle.optimizer.Optimizer` \| `None`|训练时使用的优化器。若为`None`，则使用默认定义的优化器。|`None`|
 |`save_interval_epochs`|`int`|训练时存储模型的间隔epoch数。|`1`|
 |`log_interval_steps`|`int`|训练时打印日志的间隔step数（即迭代数）。|`10`|

docs/apis/train_en.md

@@ -54,6 +54,8 @@ def train(self,
           train_dataset,
           train_batch_size=2,
           eval_dataset=None,
+          train_batch_transforms=None,
+          eval_batch_transforms=None,
           optimizer=None,
           save_interval_epochs=1,
           log_interval_steps=2,
@@ -79,6 +81,8 @@ The meaning of each parameter is as follows:
 |`train_dataset`|`paddlers.datasets.CDDataset`|Training dataset.||
 |`train_batch_size`|`int`|Batch size used during training.|`2`|
 |`eval_dataset`|`paddlers.datasets.CDDataset` \| `None`|Validation dataset.|`None`|
+|`train_batch_transforms`|`paddlers.transforms.BatchCompose` \| `None`|Batch data transformation operators in train data.|`None`|
+|`eval_batch_transforms`|`paddlers.transforms.BatchCompose` \| `None`|Batch data transformation operators in eval data.|`None`|
 |`optimizer`|`paddle.optimizer.Optimizer` \| `None`|Optimizer used during training. If `None`, the optimizer defined by default is used.|`None`|
 |`save_interval_epochs`|`int`|Number of intervals (in epochs) to evaluate and store models during training.|`1`|
 |`log_interval_steps`|`int`|Number of interval steps (i.e., the number of iterations) to print logs during training.|`2`|

docs/data/dataset_cn.md

@@ -5,7 +5,7 @@
 PaddleRS搜集汇总了遥感领域常用的**开源**深度学习数据集，提供每个数据集的以下信息：数据集说明，影像信息，标注信息，源地址，AI Studio备份链接。这些数据集按任务类型可分为**图像分类、图像分割、变化检测、目标检测、目标跟踪、多标签分类、图像生成**等多种类型。目前已收录的遥感数据集有：
 
 * 图像分类数据集32个；
-* 目标检测数据集40个；
+* 目标检测数据集41个；
 * 图像分割数据集70个；
 * 变化检测数据集28个；
 * 实例分割数据集7个；

docs/data/dataset_en.md

@@ -5,7 +5,7 @@
 PaddleRS has collected and summarized the most commonly used **open source** datasets in the field of remote sensing, providing the following information for each data set: dataset description, image information, annotation information, source address, and AI Studio backup link. According to the task type, these data sets can be divided into **image classification, image segmentation, change detection, object detection, object tracking, multi-label classification, image generation**, and other types. Currently, the collected remote sensing datasets include:
 
 * 32 image classification datasets;
-* 40 object detection datasets;
+* 41 object detection datasets;
 * 70 image segmentation datasets;
 * 28 change detection datasets;
 * 7 instance segmentation datasets;

docs/intro/data_prep_cn.md

@@ -10,4 +10,5 @@
 | 变化检测 | Season-varying | https://paperswithcode.com/dataset/cdd-dataset-season-varying | [prepare_svcd.py](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_svcd.py) |
 | 场景分类 | UC Merced | http://weegee.vision.ucmerced.edu/datasets/landuse.html | [prepare_ucmerced.py](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_ucmerced.py) |
 | 目标检测 | RSOD | https://github.com/RSIA-LIESMARS-WHU/RSOD-Dataset- | [prepare_rsod](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_rsod.py) |
+| 旋转目标检测 | DOTA | https://captain-whu.github.io/DOTA/ | [prepare_dota](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_dota.py) |
 | 图像分割 | iSAID | https://captain-whu.github.io/iSAID/ | [prepare_isaid](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_isaid.py) |

docs/intro/data_prep_en.md

@@ -10,4 +10,5 @@
 | Change Detection | Season-varying | https://paperswithcode.com/dataset/cdd-dataset-season-varying | [prepare_svcd.py](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_svcd.py) |
 | Scene Classification | UC Merced | http://weegee.vision.ucmerced.edu/datasets/landuse.html | [prepare_ucmerced.py](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_ucmerced.py) |
 | Object Detection | RSOD | https://github.com/RSIA-LIESMARS-WHU/RSOD-Dataset- | [prepare_rsod](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_rsod.py) |
+| Rotated Object Detection | DOTA | https://captain-whu.github.io/DOTA/ | [prepare_dota](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_dota.py) |
 | Image Segmentation | iSAID | https://captain-whu.github.io/iSAID/ | [prepare_isaid](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_isaid.py) |

docs/intro/model_cons_params_cn.md

@@ -449,6 +449,7 @@
 
 | 参数名 | 描述                            | 默认值 |
 | --- |-------------------------------| --- |
+| `rotate (bool)` | 表示是否是旋转框检测 | `False` |
 | `num_classes (int)` | 目标类别数量                        | `80` |
 | `backbone (str)` | 骨干网络名称                | `'MobileNetV1'` |
 | `anchors (list[list[int]])` | 预定义锚框的大小                       | `[[10, 13], [16, 30], [33, 23], [30, 61], [62, 45], [59, 119], [116, 90], [156, 198], [373, 326]]` |

docs/intro/model_cons_params_en.md

@@ -443,6 +443,7 @@ The YOLOv3 implementation based on PaddlePaddle.
 
 | Parameter Name | Description                                                                                                                 | Default Value |
 | --- |-----------------------------------------------------------------------------------------------------------------------------| --- |
+| `rotate (bool)` | If True, the model performs rotation box detection. | `False` |
 | `num_classes (int)` | Number of target classes                                                                                                    | `80` |
 | `backbone (str)` | Backbone network to use                                                                                      | `'MobileNetV1'` |
 | `anchors (list[list[int]])` | Sizes of predefined anchor boxes                                                                                                   | `[[10, 13], [16, 30], [33, 23], [30, 61], [62, 45 ], [59, 119], [116, 90], [156, 198], [373, 326]]` |

docs/intro/model_zoo_cn.md

@@ -37,6 +37,7 @@ PaddleRS目前已支持的全部模型如下（标注\*的为遥感专用模型
 | 目标检测 | PP-YOLO Tiny | 否 |
 | 目标检测 | PP-YOLOv2 | 否 |
 | 目标检测 | YOLOv3 | 否 |
+| 目标检测 | FCOS-R | 否 |
 | 图像分割 | BiSeNet V2 | 是 |
 | 图像分割 | DeepLab V3+ | 是 |
 | 图像分割 | \*FactSeg | 是 |

docs/intro/model_zoo_en.md

@@ -37,6 +37,7 @@ All models currently supported by PaddleRS are listed below (those marked \* are
 | Object Detection | PP-YOLO Tiny | No |
 | Object Detection | PP-YOLOv2 | No |
 | Object Detection | YOLOv3 | No |
+| Rotated Object Detection | FCOS-R | No |
 | Image Segmentation | BiSeNet V2 | Yes |
 | Image Segmentation | DeepLab V3+ | Yes |
 | Image Segmentation | \*FactSeg | Yes |

paddlers/datasets/base.py

@@ -37,6 +37,7 @@ def __init__(self, data_dir, label_list, transforms, num_workers, shuffle):
 
         self.num_workers = get_num_workers(num_workers)
         self.shuffle = shuffle
+        self.batch_transform = None
 
     def __getitem__(self, idx):
         sample = construct_sample_from_dict(self.file_list[idx])

paddlers/datasets/coco.py

@@ -17,7 +17,7 @@
 import os
 import os.path as osp
 import random
-from collections import OrderedDict
+from collections import OrderedDict, defaultdict
 
 import numpy as np
 
@@ -34,7 +34,7 @@ class COCODetDataset(BaseDataset):
     Args:
         data_dir (str): Root directory of the dataset.
         image_dir (str): Directory that contains the images.
-        ann_path (str): Path to COCO annotations.
+        anno_path (str): Path to COCO annotations.
         transforms (paddlers.transforms.Compose|list): Data preprocessing and data augmentation operators to apply.
         label_list (str|None, optional): Path of the file that contains the category names. Defaults to None.
         num_workers (int|str, optional): Number of processes used for data loading. If `num_workers` is 'auto',
@@ -52,7 +52,7 @@ def __init__(self,
                  image_dir,
                  anno_path,
                  transforms,
-                 label_list,
+                 label_list=None,
                  num_workers='auto',
                  shuffle=False,
                  allow_empty=False,
@@ -83,26 +83,25 @@ def __init__(self,
         self.file_list = list()
         neg_file_list = list()
         self.labels = list()
+        self.anno_path = anno_path
 
-        annotations = dict()
-        annotations['images'] = list()
-        annotations['categories'] = list()
-        annotations['annotations'] = list()
+        annotations = defaultdict(list)
 
         cname2cid = OrderedDict()
         label_id = 0
-        with open(label_list, 'r', encoding=get_encoding(label_list)) as f:
-            for line in f.readlines():
-                cname2cid[line.strip()] = label_id
-                label_id += 1
-                self.labels.append(line.strip())
-
-        for k, v in cname2cid.items():
-            annotations['categories'].append({
-                'supercategory': 'component',
-                'id': v + 1,
-                'name': k
-            })
+        if label_list:
+            with open(label_list, 'r', encoding=get_encoding(label_list)) as f:
+                for line in f.readlines():
+                    cname2cid[line.strip()] = label_id
+                    label_id += 1
+                    self.labels.append(line.strip())
+
+            for k, v in cname2cid.items():
+                annotations['categories'].append({
+                    'supercategory': 'component',
+                    'id': v + 1,
+                    'name': k
+                })
 
         anno_path = norm_path(os.path.join(self.data_dir, anno_path))
         image_dir = norm_path(os.path.join(self.data_dir, image_dir))
@@ -155,7 +154,8 @@ def __init__(self,
             gt_classes = []
             gt_bboxs = []
             gt_scores = []
-            difficults = []
+            gt_poly = []
+            difficulties = []
 
             for inst in instances:
                 # Check gt bbox
@@ -182,31 +182,43 @@ def __init__(self,
                         'area: {} x1: {}, y1: {}, x2: {}, y2: {}.'.format(
                             img_id, float(inst['area']), x1, y1, x2, y2))
 
+                if 'segmentation' in inst and inst['iscrowd']:
+                    gt_poly.append([0.0 for _ in range(8)])
+                elif 'segmentation' in inst and inst['segmentation']:
+                    if not np.array(
+                            inst['segmentation'],
+                            dtype=object).size > 0 and not self.allow_empty:
+                        continue
+                    else:
+                        gt_poly.append(inst['segmentation'])
+
                 is_crowds.append([inst['iscrowd']])
-                gt_classes.append([inst['category_id']])
+                gt_classes.append([catid2clsid[inst['category_id']]])
                 gt_bboxs.append(inst['clean_bbox'])
                 gt_scores.append([1.])
-                difficults.append([0])
-
+                difficulties.append(inst.get('difficult', 0.))
                 annotations['annotations'].append({
                     'iscrowd': inst['iscrowd'],
                     'image_id': int(inst['image_id']),
                     'bbox': inst['clean_bbox'],
                     'area': inst['area'],
                     'category_id': inst['category_id'],
                     'id': inst['id'],
-                    'difficult': 0
+                    'difficult': inst.get('difficult', 0.)
                 })
+                if gt_poly:
+                    annotations['annotations'][-1]['gt_poly'] = gt_poly[-1]
 
             label_info = {
                 'is_crowd': np.array(is_crowds),
                 'gt_class': np.array(gt_classes),
                 'gt_bbox': np.array(gt_bboxs).astype(np.float32),
                 'gt_score': np.array(gt_scores).astype(np.float32),
-                'difficult': np.array(difficults),
+                'difficult': np.array(difficulties),
+                'gt_poly': np.array(gt_poly),
             }
 
-            if label_info['gt_bbox'].size > 0:
+            if label_info['gt_bbox'].size > 0 or label_info['gt_poly'].size > 0:
                 self.file_list.append({ ** im_info, ** label_info})
                 annotations['images'].append({
                     'height': im_h,
@@ -259,13 +271,19 @@ def __getitem__(self, idx):
                 DecodeImg(to_rgb=False)(sample),
                 DecodeImg(to_rgb=False)(sample_mix)
             ])
+
         sample['trans_info'] = []
         sample, trans_info = self.transforms(sample)
         return sample, trans_info
 
     def __len__(self):
         return self.num_samples
 
+    def get_anno_path(self):
+        if self.anno_path:
+            return norm_path(os.path.join(self.data_dir, self.anno_path))
+        return None
+
     def set_epoch(self, epoch_id):
         self._epoch = epoch_id
 

paddlers/models/ppdet/core/workspace.py

@@ -67,6 +67,15 @@ def __getattr__(self, key):
             return self[key]
         raise AttributeError("object has no attribute '{}'".format(key))
 
+    def __setattr__(self, key, value):
+        self[key] = value
+
+    def copy(self):
+        new_dict = AttrDict()
+        for k, v in self.items():
+            new_dict.update({k: v})
+        return new_dict
+
 
 global_config = AttrDict()