Register Pascal VOC dataset for detection and segmentation in Detectron2
Detectron2 recognizes Pascal VOC{2007/2012} as built-in datasets. But the built-in dataset catalog for Pascal VOC provides only the bounding-box labels. This limits the use of Pascal VOC dataset to detection tasks. However, the Pascal VOC 2012 dataset does come with class and object segmentation label as well. In this post we discuss how to register VOC 2012 dataset with the segmentation labels added to the data dict. This will allow us to use the VOC dataset with Mask R-CNN models.
Load segmentation label to data dict
We define get_pascal_dict function to read the segmentation mask and bounding box labels. The segmentation mask is converted to run-length-encoding (RLE) format and add to the dataset dict record.
def get_pascal_dicts(img_dir, d='train'):
'''
Returns VOC dataset_dicts given path to "VOC2012" folder
'''
dataset_dicts = []
filename_train = img_dir / 'ImageSets/Segmentation/{}.txt'.format(d)
with open(filename_train) as f:
imlist = [line.rstrip() for line in f]
for imgidx, imname in enumerate(imlist):
filename = img_dir / f'JPEGImages/{imname}.jpg'
xmlfilename = img_dir / f'Annotations/{imname}.xml'
# Get instance segmentation mask
segmaskfilename = img_dir / f'SegmentationObject/{imname}.png'
segmask = np.array(Image.open(segmaskfilename))
record = dict()
record["file_name"] = filename.as_posix()
height, width = segmask.shape
record["height"] = height
record["width"] = width
record["image_id"] = imgidx
# Annotation info
objs = [] # list to hold annotated objects in the image
annodata = minidom.parse(xmlfilename.as_posix())
cat_obj = annodata.getElementsByTagName('object')
bboxes = annodata.getElementsByTagName('bndbox')
for idx, o in enumerate(cat_obj): # loop thru each labeled object
obj_name = o.getElementsByTagName('name')[0].firstChild.data
bbox = o.getElementsByTagName('bndbox')[0]
xmin = int(bbox.getElementsByTagName('xmin')[0].firstChild.data)
xmax = int(bbox.getElementsByTagName('xmax')[0].firstChild.data)
ymin = int(bbox.getElementsByTagName('ymin')[0].firstChild.data)
ymax = int(bbox.getElementsByTagName('ymax')[0].firstChild.data)
# Convert segmentation mask map to RLE format
seg_dict = mask.encode(np.asarray(segmask==(idx+1), order="F"))
obj = {
"bbox": [xmin, ymin, xmax, ymax],
"bbox_mode": BoxMode.XYXY_ABS,
"segmentation": seg_dict,
"category_id": name2cat[obj_name],
}
objs.append(obj)
record["annotations"] = objs
dataset_dicts.append(record)
return dataset_dicts
The get_pascal_dict() function can be used to register the VOC dataset as follows
classes = ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor']
for d in ["train", "val"]:
DatasetCatalog.register("pascal_" + d, lambda d=d: get_pascal_dicts(PATH, d))
MetadataCatalog.get("pascal_" + d).set(thing_classes=classes)
Sample images
| a | a | a |
|---|---|---|
 |
 |
 |