[transforms] Crop the whole crop_box region in crop_detection

doctr/transforms/functional/pytorch.py

@@ -113,8 +113,8 @@ def crop_detection(
     if any(val < 0 or val > 1 for val in crop_box):
         raise AssertionError("coordinates of arg `crop_box` should be relative")
     h, w = img.shape[-2:]
-    xmin, ymin = int(round(crop_box[0] * (w - 1))), int(round(crop_box[1] * (h - 1)))
-    xmax, ymax = int(round(crop_box[2] * (w - 1))), int(round(crop_box[3] * (h - 1)))
+    xmin, ymin = int(round(crop_box[0] * w)), int(round(crop_box[1] * h))
+    xmax, ymax = int(round(crop_box[2] * w)), int(round(crop_box[3] * h))
     cropped_img = F.crop(img, ymin, xmin, ymax - ymin, xmax - xmin)
     # Crop the box
     boxes = crop_boxes(boxes, crop_box if boxes.max() <= 1 else (xmin, ymin, xmax, ymax))

tests/pytorch/test_transforms_pt.py

@@ -375,7 +375,7 @@ def test_crop_detection():
     ])
     crop_box = (12 / 50, 23 / 50, 50 / 50, 50 / 50)
     c_img, c_boxes = crop_detection(img, abs_boxes, crop_box)
-    assert c_img.shape == (3, 26, 37)
+    assert c_img.shape == (3, 27, 38)
     assert c_boxes.shape == (1, 4)
     assert np.all(c_boxes == np.array([15 - 12, 0, 35 - 12, 30 - 23])[None, ...])
 
@@ -385,10 +385,21 @@ def test_crop_detection():
     ])
     crop_box = (0.24, 0.46, 1.0, 1.0)
     c_img, c_boxes = crop_detection(img, rel_boxes, crop_box)
-    assert c_img.shape == (3, 26, 37)
+    assert c_img.shape == (3, 27, 38)
     assert c_boxes.shape == (1, 4)
     assert np.abs(c_boxes - np.array([0.06 / 0.76, 0.0, 0.46 / 0.76, 0.14 / 0.54])[None, ...]).mean() < 1e-7
 
+    # The image crop covers exactly the region described by `crop_box`
+    img = torch.arange(4 * 6, dtype=torch.float32).reshape(1, 4, 6)
+    c_img, c_boxes = crop_detection(img, np.array([[0.0, 0.0, 1.0, 1.0]]), (0.0, 0.0, 1.0, 1.0))
+    assert torch.equal(c_img, img)
+    assert np.all(c_boxes == np.array([[0.0, 0.0, 1.0, 1.0]]))
+    c_img, _ = crop_detection(img, np.array([[0.0, 0.0, 1.0, 1.0]]), (0.5, 0.5, 1.0, 1.0))
+    assert torch.equal(c_img, img[:, 2:, 3:])
+    c_img, c_boxes = crop_detection(img, np.array([[0, 0, 6, 4]]), (0.5, 0.5, 1.0, 1.0))
+    assert torch.equal(c_img, img[:, 2:, 3:])
+    assert np.all(c_boxes == np.array([[0, 0, 3, 2]]))
+
     # FP16
     img = torch.ones((3, 50, 50), dtype=torch.float16)
     c_img, _ = crop_detection(img, abs_boxes, crop_box)
@@ -475,6 +486,14 @@ def test_random_crop(target):
     assert img.shape[-1] * img.shape[-2] >= 0.4 * input_t.shape[-1] * input_t.shape[-2]
     assert 0.65 <= img.shape[-2] / img.shape[-1] <= 1.6
 
+    # A crop of the whole image keeps every pixel and the mask aligned with it
+    full_cropper = RandomCrop(scale=(1.0, 1.0), ratio=(1.0, 1.0))
+    rel_target = original_target / 50
+    sample = full_cropper(Sample(image=input_t, mask=input_m, target=rel_target.copy()))
+    assert sample.image.shape == input_t.shape
+    assert sample.mask.shape == input_m.shape
+    assert np.allclose(sample.target, rel_target)
+
 
 @pytest.mark.parametrize(
     "input_dtype, input_size",