Fix video preprocessing bug in OpenCV loader

[OrangeSodahub]

Summary

This PR fixes a bug in the OpenCV video loader used for video-file inputs. Specifically, the function load_video_frames_from_video_file_using_cv2 in

sam3/sam3/model/io_utils.py

Lines 332 to 345 in 44ef224

	# 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)
	img_mean = torch.tensor(img_mean, dtype=torch.float16).view(1, 3, 1, 1)
	img_std = torch.tensor(img_std, dtype=torch.float16).view(1, 3, 1, 1)
	if not offload_video_to_cpu:
	video_tensor = video_tensor.cuda()
	img_mean = img_mean.cuda()
	img_std = img_std.cuda()
	# normalize by mean and std
	video_tensor -= img_mean
	video_tensor /= img_std
	return video_tensor, original_height, original_width

where decoded video frames were normalized without first being scaled from [0, 255] to [0, 1], even though the normalization parameters assume [0, 1] inputs. This leads to incorrectly scaled model inputs during video inference.
While in image folder loadings, / 255.0 is correctly placed:

sam3/sam3/model/io_utils.py

Line 56 in 44ef224

img_np = img_np / 255.0

And in torchcodec, too:

sam3/sam3/model/io_utils.py

Line 689 in 44ef224

frame_resized /= 255

Changes

• divide decoded OpenCV video frames by 255.0 before mean/std normalization
• convert video tensor to torch.float16 to align with other loading approaches

Validation

• Without fixing, the images tensor after init_state as below is in [-1, 509] when loading mp4 video through opencv which is incorrect, and I have also seen the unusual results
  

  sam3/sam3/model/sam3_video_inference.py

  Lines 63 to 71 in 44ef224

  	images, orig_height, orig_width = load_resource_as_video_frames(
  	resource_path=resource_path,
  	image_size=self.image_size,
  	offload_video_to_cpu=offload_video_to_cpu,
  	img_mean=self.image_mean,
  	img_std=self.image_std,
  	async_loading_frames=async_loading_frames,
  	video_loader_type=video_loader_type,
  	)

• After fixing, the images tensor is always in [-1, 1], and the segmentation results are good