Detectron2 using custom datasets

This document explains the data set API（DatasetCatalog、MetadataCatalog） How to work , And how to use them to add custom datasets .

If you want to use a custom dataset , Reuse at the same time detectron2 Data loader （data loaders）, You need ：

1. Register your dataset （ namely , tell detectron2 How to get your dataset ）.
2. （ Optional ） Register metadata for your dataset .

In order to make detectron2 Know how to get a file named “my_dataset” Data set of , The user needs to implement a function , This function is used to return the dictionary list of the dataset mentioned below , Then tell detectron2 This function ：

def my_dataset_function():
  ...
  return list[dict] in the following format

from detectron2.data import DatasetCatalog
DatasetCatalog.register("my_dataset", my_dataset_function)
# later, to access the data:
data: List[Dict] = DatasetCatalog.get("my_dataset")

ad locum , The code snippet will be named “my_dataset” The data set of is associated with the function that returns the data . If you call more than once , The function must return the same data （ In the same order ）. Registration is always valid , Until the process exits .
This function can do anything , And should return list[dict] Data in , Every dict All in one of the following formats ：

• Detectron2 Standard dataset Dictionary , As follows . This will make it similar to detectron2 Many of the other built-in features in , Therefore, it is recommended to use it when sufficient .
• Any custom format . You can also return any... In your own format dicts, For example, add additional keys for new tasks . then , You also need to handle them correctly downstream . See more details below .
  Standard dataset dictionary
  For standard tasks （ Instance detection 、 example / semantics / Panoramic segmentation 、 Key point detection ）, We load the original dataset into the dictionary list （ list[dict]） in , Its specification is similar to COCO The annotation .
  Every dict Contains information about an image . dict May have the following fields , The required fields vary depending on the needs of the data loader or task （ See below ）

• file_name： The full path of the image file .
• height, width： Integers , The shape of the image .
• image_id（str or int）： A unique... That identifies this image id.
• annotations （list[dict]）： Instance detection / Required for segmentation or key detection tasks . Every dict The corresponding annotation of an instance in this image , And may include the following keys：

 1. bbox (list[float], required):  Represents the instance bounding box  4  A list of numbers .
 2. bbox_mode (int, required): bbox  The format of . It has to be structures.BoxMode Members of . At present, we support ：BoxMode.XYXY_ABS、BoxMode.XYWH_ABS.
 3. category_id (int, required)：[0, num_categories-1]  Range of integers , Indicates a category label . If applicable , Reserved values  num_categories  To indicate that “ background ” Category .
 4. segmentation （list[list[float]] or dict）： Split mask of the instance .
 		a.  If it is  list[list[float]], It represents a list of polygons , One for each connected component of the object . Every  list[float]  Is a simple polygon , The format is  [x1, y1, ..., xn, yn] (n≥3). Xs  and  Ys  Is the absolute coordinate in pixels .		
		b.  If it is  dict, said  COCO  Compress  RLE  Per pixel segmentation mask in format . dict  There should be a key “size” and “counts”. You can  pycocotools.mask.encode(np.asarray(mask, order="F"))  take  0s  and  1s  Of  uint8  The segmentation mask is converted to such  dict. If you use the default data loader in this format ,cfg.INPUT.MASK_FORMAT  Must be set to bitmask .		
 5. keypoints (list[float])： The format is  [x1, y1, v1,..., xn, yn, vn]. v[i]  Indicates the visibility of the key . n  Must be equal to the number of key categories . Xs  and  Ys  yes  [0, W  or  H]  Absolute real value coordinates within the range .		
		（ Be careful ,COCO  The key coordinates of the format are  [0, W-1  or  H-1]  Range of integers , This is different from our standard format .Detectron2  take  COCO  Key coordinates plus  0.5  To convert them from discrete pixel indexes to floating point coordinates .）		
 6. iscrowd：0（ Default ） or 1. Whether this instance is marked as COCO Of “ Crowd area ”. If you don't know what it means , Please do not include this field .

If annotations It's an empty list , It means that the image is marked as having no object . By default , Such images will be deleted from the training , But you can use DATALOADER.FILTER_EMPTY_ANNOTATIONS Included .

• sem_seg_file_name (str)： Semantic segmentation ground truth The full path to the file . It should be a grayscale image , Its pixel value is an integer label .
• pan_seg_file_name (str)： Panoramic segmentation ground truth The full path to the file . It should be a RGB Images , Its pixel value is using panopticapi.utils.id2rgb Function encoded integer id. id from segments_info Definition . If id
  Not in segments_info in , Then the pixel is considered unmarked , And often overlooked in training and evaluation .
• Segments_info(list[dict])： Define panoramic segmentation ground truth Each of them id The meaning of . Every dict All have the following keys ：

 1. id (int)： Appear in the ground truth Integer in image .
 2. category_id(int)：[0, num_categories-1]  Range of integers , Indicates a category label .
 3. iscrowd：0（ Default ） or 1. Whether this instance is marked as COCO Of “ Crowd area ”.