‍‍‍‍CocoDetection accepts slice as argument to its __getitem__ and returns wrong result

[sadra-barikbin]

Hi, although CocoDetection declares its __getitem__ argument to be of type int, it actually accepts slice as well. However it returns a wrong result. As this might be confusing for some users, I suggest to add a type check at the beginning of the __getitem__.

vision/torchvision/datasets/coco.py

Lines 46 to 49 in cab01fc

	def __getitem__(self, index: int) -> Tuple[Any, Any]:
	id = self.ids[index]
	image = self._load_image(id)
	target = self._load_target(id)

ds[:2]

(tensor([[[0.3373, 0.3373, 0.3373,  ..., 0.2863, 0.2863, 0.2863],
          [0.3373, 0.3373, 0.3373,  ..., 0.2863, 0.2863, 0.2863],
          [0.3373, 0.3373, 0.3373,  ..., 0.2863, 0.2863, 0.2863],
          ...,
          [0.4196, 0.4196, 0.4196,  ..., 0.4137, 0.3980, 0.3235],
          [0.4196, 0.4196, 0.4196,  ..., 0.4137, 0.3980, 0.3235],
          [0.4196, 0.4196, 0.4196,  ..., 0.4137, 0.3980, 0.3235]],
 
         [[0.1608, 0.1608, 0.1608,  ..., 0.1216, 0.1216, 0.1216],
          [0.1608, 0.1608, 0.1608,  ..., 0.1216, 0.1216, 0.1216],
          [0.1608, 0.1608, 0.1608,  ..., 0.1216, 0.1216, 0.1216],
          ...,
          [0.5686, 0.5686, 0.5686,  ..., 0.5333, 0.5137, 0.4353],
          [0.5686, 0.5686, 0.5686,  ..., 0.5333, 0.5137, 0.4353],
          [0.5686, 0.5686, 0.5686,  ..., 0.5333, 0.5137, 0.4353]],
 
         [[0.1412, 0.1412, 0.1412,  ..., 0.1608, 0.1608, 0.1608],
          [0.1412, 0.1412, 0.1412,  ..., 0.1608, 0.1608, 0.1608],
          [0.1412, 0.1412, 0.1412,  ..., 0.1608, 0.1608, 0.1608],
          ...,
          [0.0902, 0.0902, 0.0902,  ..., 0.1333, 0.1451, 0.1020],
          [0.0902, 0.0902, 0.0902,  ..., 0.1333, 0.1451, 0.1020],
          [0.0902, 0.0902, 0.0902,  ..., 0.1333, 0.1451, 0.1020]]]),
 {'category_id': tensor([2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 3,
          1]),
  'bbox': tensor([[ 390,  240,   58,   92],
          [ 391,  319,   66,   99],
          [ 663,  331,   54,  106],
          [ 796,  324,   50,  102],
          [1245,  276,   35,  117],
          [ 599,  265,   58,   91],
          [ 755,  228,   42,   91],
          [ 788,  261,   61,   97],
          [ 250,  374,   73,  105],
          [ 224,  449,   29,   28],
          [  12,   85,   13,   32],
          [  88,   70,   13,   27],
          [ 102,   99,   14,   33],
          [ 100,  132,   14,   32],
          [ 151,  160,   13,   34],
          [ 177,  115,   18,   32],
          [ 214,   73,   12,   28],
          [ 267,   58,   13,   23],
          [ 358,   71,   14,   33],
          [ 356,  160,   17,   36],
          [ 231,  132,   14,   33],
          [ 234,   88,   16,   29],
          [ 248,   90,   13,   25],
          [  60,  109,   13,   32],
          [ 243,   75,    7,    5]]),
  'iscrowd': tensor([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
          0])})

The image in the returned tuple is that of the ds[0] and the target is the combined annotations of ds[0] and ds[1] which is wrong. This is due to the fact that ds.coco could work with a list of inputs in its getAnnIds().

cc @pmeier