Train Faster R-CNN with negative samples (#1911)

* modified FasterRCNN to accept negative samples

* remove debug lines

* Change torch.zeros_like to torch.zerros

* Add unit tests

* take the `device` into account
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

test/test_models_detection_negative_samples.py

+import torch
+
+import torchvision.models
+from torchvision.ops import MultiScaleRoIAlign
+from torchvision.models.detection.rpn import AnchorGenerator, RPNHead, RegionProposalNetwork
+from torchvision.models.detection.roi_heads import RoIHeads
+from torchvision.models.detection.faster_rcnn import FastRCNNPredictor, TwoMLPHead
+
+import unittest
+
+
+class Tester(unittest.TestCase):
+
+    def test_targets_to_anchors(self):
+        boxes = torch.zeros((0, 4), dtype=torch.float32)
+        negative_target = {"boxes": boxes,
+                           "labels": torch.zeros((1, 1), dtype=torch.int64),
+                           "image_id": 4,
+                           "area": (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0]),
+                           "iscrowd": torch.zeros((0,), dtype=torch.int64)}
+
+        anchors = [torch.randint(-50, 50, (3, 4), dtype=torch.float32)]
+        targets = [negative_target]
+
+        anchor_sizes = ((32,), (64,), (128,), (256,), (512,))
+        aspect_ratios = ((0.5, 1.0, 2.0),) * len(anchor_sizes)
+        rpn_anchor_generator = AnchorGenerator(
+            anchor_sizes, aspect_ratios
+        )
+        rpn_head = RPNHead(4, rpn_anchor_generator.num_anchors_per_location()[0])
+
+        head = RegionProposalNetwork(
+            rpn_anchor_generator, rpn_head,
+            0.5, 0.3,
+            256, 0.5,
+            2000, 2000, 0.7)
+
+        labels, matched_gt_boxes = head.assign_targets_to_anchors(anchors, targets)
+
+        self.assertEqual(labels[0].sum(), 0)
+        self.assertEqual(labels[0].shape, torch.Size([anchors[0].shape[0]]))
+        self.assertEqual(labels[0].dtype, torch.float32)
+
+        self.assertEqual(matched_gt_boxes[0].sum(), 0)
+        self.assertEqual(matched_gt_boxes[0].shape, anchors[0].shape)
+        self.assertEqual(matched_gt_boxes[0].dtype, torch.float32)
+
+    def test_assign_targets_to_proposals(self):
+
+        proposals = [torch.randint(-50, 50, (20, 4), dtype=torch.float32)]
+        gt_boxes = [torch.zeros((0, 4), dtype=torch.float32)]
+        gt_labels = [torch.tensor([[0]], dtype=torch.int64)]
+
+        box_roi_pool = MultiScaleRoIAlign(
+            featmap_names=['0', '1', '2', '3'],
+            output_size=7,
+            sampling_ratio=2)
+
+        resolution = box_roi_pool.output_size[0]
+        representation_size = 1024
+        box_head = TwoMLPHead(
+            4 * resolution ** 2,
+            representation_size)
+
+        representation_size = 1024
+        box_predictor = FastRCNNPredictor(
+            representation_size,
+            2)
+
+        roi_heads = RoIHeads(
+            # Box
+            box_roi_pool, box_head, box_predictor,
+            0.5, 0.5,
+            512, 0.25,
+            None,
+            0.05, 0.5, 100)
+
+        matched_idxs, labels = roi_heads.assign_targets_to_proposals(proposals, gt_boxes, gt_labels)
+
+        self.assertEqual(matched_idxs[0].sum(), 0)
+        self.assertEqual(matched_idxs[0].shape, torch.Size([proposals[0].shape[0]]))
+        self.assertEqual(matched_idxs[0].dtype, torch.int64)
+
+        self.assertEqual(labels[0].sum(), 0)
+        self.assertEqual(labels[0].shape, torch.Size([proposals[0].shape[0]]))
+        self.assertEqual(labels[0].dtype, torch.int64)
+
+    def test_forward_negative_sample(self):
+        model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
+        in_features = model.roi_heads.box_predictor.cls_score.in_features
+        model.roi_heads.box_predictor = FastRCNNPredictor(in_features, 2)
+
+        images = [torch.rand((3, 100, 100), dtype=torch.float32)]
+        boxes = torch.zeros((0, 4), dtype=torch.float32)
+        negative_target = {"boxes": boxes,
+                           "labels": torch.zeros((1, 1), dtype=torch.int64),
+                           "image_id": 4,
+                           "area": (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0]),
+                           "iscrowd": torch.zeros((0,), dtype=torch.int64)}
+
+        targets = [negative_target]
+
+        loss_dict = model(images, targets)
+
+        self.assertEqual(loss_dict["loss_box_reg"], torch.tensor(0.))
+        self.assertEqual(loss_dict["loss_rpn_box_reg"], torch.tensor(0.))
+
+
+if __name__ == '__main__':
+    unittest.main()

torchvision/models/detection/roi_heads.py

@@ -574,6 +574,17 @@ class RoIHeads(torch.nn.Module):
         matched_idxs = []
         labels = []
         for proposals_in_image, gt_boxes_in_image, gt_labels_in_image in zip(proposals, gt_boxes, gt_labels):
+
+            if gt_boxes_in_image.numel() == 0:
+                # Background image
+                device = proposals_in_image.device
+                clamped_matched_idxs_in_image = torch.zeros(
+                    (proposals_in_image.shape[0],), dtype=torch.int64, device=device
+                )
+                labels_in_image = torch.zeros(
+                    (proposals_in_image.shape[0],), dtype=torch.int64, device=device
+                )
+            else:
                 #  set to self.box_similarity when https://github.com/pytorch/pytorch/issues/27495 lands
                 match_quality_matrix = box_ops.box_iou(gt_boxes_in_image, proposals_in_image)
                 matched_idxs_in_image = self.proposal_matcher(match_quality_matrix)
@@ -635,6 +646,8 @@ class RoIHeads(torch.nn.Module):
         self.check_targets(targets)
         assert targets is not None
         dtype = proposals[0].dtype
+        device = proposals[0].device
+
         gt_boxes = [t["boxes"].to(dtype) for t in targets]
         gt_labels = [t["labels"] for t in targets]
 
@@ -652,7 +665,11 @@ class RoIHeads(torch.nn.Module):
             proposals[img_id] = proposals[img_id][img_sampled_inds]
             labels[img_id] = labels[img_id][img_sampled_inds]
             matched_idxs[img_id] = matched_idxs[img_id][img_sampled_inds]
-            matched_gt_boxes.append(gt_boxes[img_id][matched_idxs[img_id]])
+
+            gt_boxes_in_image = gt_boxes[img_id]
+            if gt_boxes_in_image.numel() == 0:
+                gt_boxes_in_image = torch.zeros((1, 4), dtype=dtype, device=device)
+            matched_gt_boxes.append(gt_boxes_in_image[matched_idxs[img_id]])
 
         regression_targets = self.box_coder.encode(matched_gt_boxes, proposals)
         return proposals, matched_idxs, labels, regression_targets

torchvision/models/detection/rpn.py

@@ -336,6 +336,13 @@ class RegionProposalNetwork(torch.nn.Module):
         matched_gt_boxes = []
         for anchors_per_image, targets_per_image in zip(anchors, targets):
             gt_boxes = targets_per_image["boxes"]
+
+            if gt_boxes.numel() == 0:
+                # Background image (negative example)
+                device = anchors_per_image.device
+                matched_gt_boxes_per_image = torch.zeros(anchors_per_image.shape, dtype=torch.float32, device=device)
+                labels_per_image = torch.zeros((anchors_per_image.shape[0],), dtype=torch.float32, device=device)
+            else:
                 match_quality_matrix = box_ops.box_iou(gt_boxes, anchors_per_image)
                 matched_idxs = self.proposal_matcher(match_quality_matrix)
                 # get the targets corresponding GT for each proposal