Fix OpenCV video loader normalization scale

Author: SakshamKapoor2911

sam3/model/io_utils.py

@@ -342,8 +342,9 @@ def load_video_frames_from_video_file_using_cv2(
         )
 
     # Convert to tensor
-    frames_np = np.stack(frames, axis=0).astype(np.float32)  # (T, H, W, C)
-    video_tensor = torch.from_numpy(frames_np).permute(0, 3, 1, 2)  # (T, C, H, W)
+    frames_np = np.stack(frames, axis=0)  # (T, H, W, C)
+    video_tensor = torch.from_numpy(frames_np).permute(0, 3, 1, 2)
+    video_tensor = video_tensor.to(dtype=torch.float16)  # (T, C, H, W)
 
     # pyrefly: ignore [bad-assignment]
     img_mean = torch.tensor(img_mean, dtype=torch.float16).view(1, 3, 1, 1)
@@ -357,6 +358,7 @@ def load_video_frames_from_video_file_using_cv2(
         img_std = img_std.cuda()
     # normalize by mean and std
     # pyrefly: ignore [unsupported-operation]
+    video_tensor /= 255.0
     video_tensor -= img_mean
     # pyrefly: ignore [unsupported-operation]
     video_tensor /= img_std

test/test_io_utils.py

@@ -2,10 +2,14 @@
 
 """Tests for io_utils extensionless video file handling (D99228861)."""
 
+import os
 import tempfile
 import unittest
 from unittest.mock import MagicMock, patch
 
+import numpy as np
+import torch
+
 from sam3.model.io_utils import load_video_frames
 
 
@@ -107,6 +111,89 @@ def test_dummy_video_pattern(self) -> None:
         self.assertEqual(h, 480)
         self.assertEqual(w, 640)
 
+    def test_cv2_video_file_loader_scales_before_normalization(self) -> None:
+        """OpenCV video loading should match normalized decoded uint8 frames."""
+        try:
+            import cv2
+        except ImportError as exc:
+            self.skipTest(f"OpenCV is required for this test: {exc}")
+
+        image_size = 8
+        source_height = 6
+        source_width = 10
+        img_mean = (0.5, 0.25, 0.75)
+        img_std = (0.5, 0.25, 0.25)
+        yy, xx = np.indices((source_height, source_width), dtype=np.uint16)
+        frames_rgb = [
+            np.stack(
+                (
+                    (xx * 23 + yy * 7) % 256,
+                    (xx * 11 + yy * 17 + 3) % 256,
+                    (xx * 5 + yy * 29 + 9) % 256,
+                ),
+                axis=-1,
+            ).astype(np.uint8),
+            np.stack(
+                (
+                    (xx * 13 + yy * 19 + 31) % 256,
+                    (xx * 3 + yy * 41 + 47) % 256,
+                    (xx * 37 + yy * 2 + 61) % 256,
+                ),
+                axis=-1,
+            ).astype(np.uint8),
+        ]
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            video_path = os.path.join(tmpdir, "tiny.avi")
+            writer = cv2.VideoWriter(
+                video_path,
+                cv2.VideoWriter_fourcc(*"MJPG"),
+                2.0,
+                (source_width, source_height),
+            )
+            if not writer.isOpened():
+                self.skipTest("OpenCV could not create a temporary MJPG video")
+
+            for frame_rgb in frames_rgb:
+                writer.write(cv2.cvtColor(frame_rgb, cv2.COLOR_RGB2BGR))
+            writer.release()
+
+            decoded_frames = []
+            cap = cv2.VideoCapture(video_path)
+            while True:
+                ret, frame_bgr = cap.read()
+                if not ret:
+                    break
+                frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
+                decoded_frames.append(
+                    cv2.resize(
+                        frame_rgb,
+                        (image_size, image_size),
+                        interpolation=cv2.INTER_CUBIC,
+                    )
+                )
+            cap.release()
+            self.assertEqual(len(decoded_frames), len(frames_rgb))
+
+            expected = torch.from_numpy(np.stack(decoded_frames, axis=0))
+            expected = expected.permute(0, 3, 1, 2).to(dtype=torch.float16)
+            expected /= 255.0
+            expected -= torch.tensor(img_mean, dtype=torch.float16).view(1, 3, 1, 1)
+            expected /= torch.tensor(img_std, dtype=torch.float16).view(1, 3, 1, 1)
+
+            frames, height, width = load_video_frames(
+                video_path=video_path,
+                image_size=image_size,
+                offload_video_to_cpu=True,
+                img_mean=img_mean,
+                img_std=img_std,
+                video_loader_type="cv2",
+            )
+
+        self.assertEqual((height, width), (source_height, source_width))
+        self.assertEqual(frames.dtype, torch.float16)
+        torch.testing.assert_close(frames, expected, rtol=0, atol=1e-6)
+
     @patch("sam3.model.io_utils.load_video_frames_from_video_file")
     def test_unknown_extension_routes_to_video_loader(
         self, mock_load_video: MagicMock