Fix empty anno - merge back develop (#4022)

Refactor empty label workaround in iseg and mask_target.py

Author: eugene123tw

src/otx/algo/common/utils/bbox_overlaps.py

@@ -8,6 +8,8 @@
 
 from __future__ import annotations
 
+import warnings
+
 import torch
 from torch import Tensor
 
@@ -142,15 +144,28 @@ def bbox_overlaps(
         >>> assert tuple(bbox_overlaps(nonempty, empty).shape) == (1, 0)
         >>> assert tuple(bbox_overlaps(empty, empty).shape) == (0, 0)
     """
+    if not (bboxes1.size(-1) == 4 or bboxes1.size(0) == 0):
+        msg = "bboxes1 must have a last dimension of size 4 or be an empty tensor."
+        raise ValueError(msg)
+
+    if not (bboxes2.size(-1) == 4 or bboxes2.size(0) == 0):
+        msg = "bboxes2 must have a last dimension of size 4 or be an empty tensor."
+        raise ValueError(msg)
+
+    if bboxes1.shape[:-2] != bboxes2.shape[:-2]:
+        msg = "The batch dimension of bboxes must be the same."
+        raise ValueError(msg)
+
     batch_shape = bboxes1.shape[:-2]
 
     rows = bboxes1.size(-2)
     cols = bboxes2.size(-2)
 
     if rows * cols == 0:
+        warnings.warn("No bboxes are provided! Returning empty boxes!", stacklevel=2)
         if is_aligned:
-            return bboxes1.new((*batch_shape, rows))
-        return bboxes1.new((*batch_shape, rows, cols))
+            return bboxes1.new(batch_shape + (rows,))  # noqa: RUF005
+        return bboxes1.new(batch_shape + (rows, cols))  # noqa: RUF005
 
     area1 = (bboxes1[..., 2] - bboxes1[..., 0]) * (bboxes1[..., 3] - bboxes1[..., 1])
     area2 = (bboxes2[..., 2] - bboxes2[..., 0]) * (bboxes2[..., 3] - bboxes2[..., 1])

src/otx/algo/instance_segmentation/utils/structures/mask/mask_target.py

@@ -8,6 +8,8 @@
 
 from __future__ import annotations
 
+import warnings
+
 import numpy as np
 import torch
 from datumaro.components.annotation import Polygon
@@ -62,16 +64,20 @@ def mask_target_single(
     meta_info: dict,
 ) -> Tensor:
     """Compute mask target for each positive proposal in the image."""
+    mask_size = _pair(mask_size)
+    if len(gt_masks) == 0:
+        warnings.warn("No ground truth masks are provided!", stacklevel=2)
+        return pos_proposals.new_zeros((0, *mask_size))
+
     if isinstance(gt_masks[0], Polygon):
         crop_and_resize = crop_and_resize_polygons
     elif isinstance(gt_masks, tv_tensors.Mask):
         crop_and_resize = crop_and_resize_masks
     else:
-        msg = f"Unsupported type of masks: {type(gt_masks[0])}"
-        raise NotImplementedError(msg)
+        warnings.warn("Unsupported ground truth mask type!", stacklevel=2)
+        return pos_proposals.new_zeros((0, *mask_size))
 
     device = pos_proposals.device
-    mask_size = _pair(mask_size)
     num_pos = pos_proposals.size(0)
     if num_pos > 0:
         proposals_np = pos_proposals.cpu().numpy()

src/otx/core/data/pre_filtering.py

@@ -40,13 +40,15 @@ def pre_filtering(
     dataset = DmDataset.filter(dataset, is_valid_annot, filter_annotations=True)
     dataset = remove_unused_labels(dataset, data_format, ignore_index)
     if unannotated_items_ratio > 0:
-        empty_items = [item.id for item in dataset if item.subset == "train" and len(item.annotations) == 0]
+        empty_items = [
+            item.id for item in dataset if item.subset in ("train", "TRAINING") and len(item.annotations) == 0
+        ]
         used_background_items = set(sample(empty_items, int(len(empty_items) * unannotated_items_ratio)))
 
     return DmDataset.filter(
         dataset,
         lambda item: not (
-            item.subset == "train" and len(item.annotations) == 0 and item.id not in used_background_items
+            item.subset in ("train", "TRAINING") and len(item.annotations) == 0 and item.id not in used_background_items
         ),
     )
 

tests/unit/algo/common/test_iou2d_calculator.py

@@ -0,0 +1,123 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+# Copyright (c) OpenMMLab. All rights reserved.
+from __future__ import annotations
+
+import numpy as np
+import pytest
+import torch
+from otx.algo.common.utils.assigners.iou2d_calculator import BboxOverlaps2D
+from otx.algo.common.utils.bbox_overlaps import bbox_overlaps
+
+
+def test_bbox_overlaps_2d(eps: float = 1e-7):
+    def _construct_bbox(num_bbox: int | None = None) -> tuple[torch.Tensor, int]:
+        img_h = int(np.random.randint(3, 1000))
+        img_w = int(np.random.randint(3, 1000))
+        if num_bbox is None:
+            num_bbox = np.random.randint(1, 10)
+        x1y1 = torch.rand((num_bbox, 2))
+        x2y2 = torch.max(torch.rand((num_bbox, 2)), x1y1)
+        bboxes = torch.cat((x1y1, x2y2), -1)
+        bboxes[:, 0::2] *= img_w
+        bboxes[:, 1::2] *= img_h
+        return bboxes, num_bbox
+
+    # Test where is_aligned is True, bboxes.size(-1) == 5 (include score)
+    self = BboxOverlaps2D()
+    bboxes1, num_bbox = _construct_bbox()
+    bboxes2, _ = _construct_bbox(num_bbox)
+    bboxes1 = torch.cat((bboxes1, torch.rand((num_bbox, 1))), 1)
+    bboxes2 = torch.cat((bboxes2, torch.rand((num_bbox, 1))), 1)
+    gious = self(bboxes1, bboxes2, "giou", True)
+    assert gious.size() == (num_bbox,), gious.size()
+    assert torch.all(gious >= -1)
+    assert torch.all(gious <= 1)
+
+    # Test where is_aligned is True, bboxes1.size(-2) == 0
+    bboxes1 = torch.empty((0, 4))
+    bboxes2 = torch.empty((0, 4))
+    gious = self(bboxes1, bboxes2, "giou", True)
+    assert gious.size() == (0,), gious.size()
+    assert torch.all(gious == torch.empty((0,)))
+    assert torch.all(gious >= -1)
+    assert torch.all(gious <= 1)
+
+    # Test where is_aligned is True, and bboxes.ndims > 2
+    bboxes1, num_bbox = _construct_bbox()
+    bboxes2, _ = _construct_bbox(num_bbox)
+    bboxes1 = bboxes1.unsqueeze(0).repeat(2, 1, 1)
+    # test assertion when batch dim is not the same
+    with pytest.raises(ValueError, match="The batch dimension of bboxes must be the same."):
+        self(bboxes1, bboxes2.unsqueeze(0).repeat(3, 1, 1), "giou", True)
+    bboxes2 = bboxes2.unsqueeze(0).repeat(2, 1, 1)
+    gious = self(bboxes1, bboxes2, "giou", True)
+    assert torch.all(gious >= -1)
+    assert torch.all(gious <= 1)
+    assert gious.size() == (2, num_bbox)
+    bboxes1 = bboxes1.unsqueeze(0).repeat(2, 1, 1, 1)
+    bboxes2 = bboxes2.unsqueeze(0).repeat(2, 1, 1, 1)
+    gious = self(bboxes1, bboxes2, "giou", True)
+    assert torch.all(gious >= -1)
+    assert torch.all(gious <= 1)
+    assert gious.size() == (2, 2, num_bbox)
+
+    # Test where is_aligned is False
+    bboxes1, num_bbox1 = _construct_bbox()
+    bboxes2, num_bbox2 = _construct_bbox()
+    gious = self(bboxes1, bboxes2, "giou")
+    assert torch.all(gious >= -1)
+    assert torch.all(gious <= 1)
+    assert gious.size() == (num_bbox1, num_bbox2)
+
+    # Test where is_aligned is False, and bboxes.ndims > 2
+    bboxes1 = bboxes1.unsqueeze(0).repeat(2, 1, 1)
+    bboxes2 = bboxes2.unsqueeze(0).repeat(2, 1, 1)
+    gious = self(bboxes1, bboxes2, "giou")
+    assert torch.all(gious >= -1)
+    assert torch.all(gious <= 1)
+    assert gious.size() == (2, num_bbox1, num_bbox2)
+    bboxes1 = bboxes1.unsqueeze(0)
+    bboxes2 = bboxes2.unsqueeze(0)
+    gious = self(bboxes1, bboxes2, "giou")
+    assert torch.all(gious >= -1)
+    assert torch.all(gious <= 1)
+    assert gious.size() == (1, 2, num_bbox1, num_bbox2)
+
+    # Test where is_aligned is False, bboxes1.size(-2) == 0
+    gious = self(torch.empty(1, 2, 0, 4), bboxes2, "giou")
+    assert torch.all(gious == torch.empty(1, 2, 0, bboxes2.size(-2)))
+    assert torch.all(gious >= -1)
+    assert torch.all(gious <= 1)
+
+    # test allclose between bbox_overlaps and the original official
+    # implementation.
+    bboxes1 = torch.FloatTensor(
+        [
+            [0, 0, 10, 10],
+            [10, 10, 20, 20],
+            [32, 32, 38, 42],
+        ],
+    )
+    bboxes2 = torch.FloatTensor(
+        [
+            [0, 0, 10, 20],
+            [0, 10, 10, 19],
+            [10, 10, 20, 20],
+        ],
+    )
+    gious = bbox_overlaps(bboxes1, bboxes2, "giou", is_aligned=True, eps=eps)
+    gious = gious.numpy().round(4)
+    # the gt is got with four decimal precision.
+    expected_gious = np.array([0.5000, -0.0500, -0.8214])
+    assert np.allclose(gious, expected_gious, rtol=0, atol=eps)
+
+    # test mode 'iof'
+    ious = bbox_overlaps(bboxes1, bboxes2, "iof", is_aligned=True, eps=eps)
+    assert torch.all(ious >= -1)
+    assert torch.all(ious <= 1)
+    assert ious.size() == (bboxes1.size(0),)
+    ious = bbox_overlaps(bboxes1, bboxes2, "iof", eps=eps)
+    assert torch.all(ious >= -1)
+    assert torch.all(ious <= 1)
+    assert ious.size() == (bboxes1.size(0), bboxes2.size(0))