[Bugfix] Align OpenCV video metadata timeline (#47099)
Signed-off-by: VectorPeak <73048950+VectorPeak@users.noreply.github.com>
This commit is contained in:
@@ -10,7 +10,11 @@ import pytest
|
||||
from PIL import Image
|
||||
|
||||
from vllm.assets.base import get_vllm_public_assets
|
||||
from vllm.assets.video import video_to_ndarrays, video_to_pil_images_list
|
||||
from vllm.assets.video import (
|
||||
video_get_metadata,
|
||||
video_to_ndarrays,
|
||||
video_to_pil_images_list,
|
||||
)
|
||||
from vllm.multimodal.media import ImageMediaIO, VideoMediaIO
|
||||
from vllm.multimodal.video import VIDEO_LOADER_REGISTRY, VideoLoader
|
||||
|
||||
@@ -112,6 +116,20 @@ def test_opencv_video_io_colorspace(tmp_path, is_color: bool, fourcc: str, ext:
|
||||
assert np.nanmean(sim) > 0.99
|
||||
|
||||
|
||||
def test_opencv_video_metadata_matches_sampled_frame_timeline(tmp_path):
|
||||
image_path = f"{tmp_path}/test_metadata_image.png"
|
||||
Image.new("RGB", (8, 8), color=(255, 0, 0)).save(image_path)
|
||||
video_path = f"{tmp_path}/test_metadata_video.mp4"
|
||||
create_video_from_image(image_path, video_path, num_frames=10, fps=5.0)
|
||||
|
||||
metadata = video_get_metadata(video_path, num_frames=4)
|
||||
|
||||
assert metadata["fps"] == pytest.approx(5.0)
|
||||
assert metadata["duration"] == pytest.approx(2.0)
|
||||
assert metadata["frames_indices"] == [0, 3, 6, 9]
|
||||
assert metadata["total_num_frames"] == 4
|
||||
|
||||
|
||||
NUM_FRAMES = 10
|
||||
FAKE_OUTPUT_1 = np.random.rand(NUM_FRAMES, 1280, 720, 3)
|
||||
FAKE_OUTPUT_2 = np.random.rand(NUM_FRAMES, 1280, 720, 3)
|
||||
|
||||
@@ -15,6 +15,10 @@ from vllm.transformers_utils.repo_utils import hf_api
|
||||
from .base import get_cache_dir
|
||||
|
||||
|
||||
def _sample_frame_indices(total_frames: int, num_frames: int) -> npt.NDArray:
|
||||
return np.linspace(0, total_frames - 1, num_frames, dtype=int)
|
||||
|
||||
|
||||
@lru_cache
|
||||
def download_video_asset(filename: str) -> str:
|
||||
"""
|
||||
@@ -47,7 +51,7 @@ def video_to_ndarrays(path: str, num_frames: int = -1) -> npt.NDArray:
|
||||
frames = []
|
||||
|
||||
num_frames = num_frames if num_frames > 0 else total_frames
|
||||
frame_indices = np.linspace(0, total_frames - 1, num_frames, dtype=int)
|
||||
frame_indices = _sample_frame_indices(total_frames, num_frames)
|
||||
for idx in range(total_frames):
|
||||
ok = cap.grab() # next img
|
||||
if not ok:
|
||||
@@ -86,13 +90,14 @@ def video_get_metadata(path: str, num_frames: int = -1) -> dict[str, Any]:
|
||||
|
||||
if num_frames == -1 or num_frames > total_frames:
|
||||
num_frames = total_frames
|
||||
frame_indices = _sample_frame_indices(total_frames, num_frames)
|
||||
|
||||
metadata = {
|
||||
"total_num_frames": num_frames,
|
||||
"fps": duration / num_frames,
|
||||
"fps": fps,
|
||||
"duration": duration,
|
||||
"video_backend": "opencv",
|
||||
"frames_indices": list(range(num_frames)),
|
||||
"frames_indices": frame_indices.tolist(),
|
||||
# extra field used to control hf processor's video
|
||||
# sampling behavior
|
||||
"do_sample_frames": num_frames == total_frames,
|
||||
|
||||
Reference in New Issue
Block a user