sam3/model/utils/sam2_utils.py

# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import os
from threading import Thread

import numpy as np
import torch
from PIL import Image
from tqdm import tqdm


def _load_img_as_tensor(img_path, image_size):
    img_pil = Image.open(img_path)
    img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
    if img_np.dtype == np.uint8:  # np.uint8 is expected for JPEG images
        img_np = img_np / 255.0
    else:
        raise RuntimeError(f"Unknown image dtype: {img_np.dtype} on {img_path}")
    img = torch.from_numpy(img_np).permute(2, 0, 1)
    video_width, video_height = img_pil.size  # the original video size
    return img, video_height, video_width


class AsyncVideoFrameLoader:
    """
    A list of video frames to be load asynchronously without blocking session start.
    """

    def __init__(
        self,
        img_paths,
        image_size,
        offload_video_to_cpu,
        img_mean,
        img_std,
        compute_device,
    ):
        self.img_paths = img_paths
        self.image_size = image_size
        self.offload_video_to_cpu = offload_video_to_cpu
        self.img_mean = img_mean
        self.img_std = img_std
        # items in `self.images` will be loaded asynchronously
        self.images = [None] * len(img_paths)
        # catch and raise any exceptions in the async loading thread
        self.exception = None
        # video_height and video_width be filled when loading the first image
        self.video_height = None
        self.video_width = None
        self.compute_device = compute_device

        # load the first frame to fill video_height and video_width and also
        # to cache it (since it's most likely where the user will click)
        self.__getitem__(0)

        # load the rest of frames asynchronously without blocking the session start
        def _load_frames():
            try:
                for n in tqdm(range(len(self.images)), desc="frame loading (JPEG)"):
                    self.__getitem__(n)
            except Exception as e:
                self.exception = e

        self.thread = Thread(target=_load_frames, daemon=True)
        self.thread.start()

    def __getitem__(self, index):
        if self.exception is not None:
            raise RuntimeError("Failure in frame loading thread") from self.exception

        img = self.images[index]
        if img is not None:
            return img

        img, video_height, video_width = _load_img_as_tensor(
            self.img_paths[index], self.image_size
        )
        self.video_height = video_height
        self.video_width = video_width
        # normalize by mean and std
        img -= self.img_mean
        img /= self.img_std
        if not self.offload_video_to_cpu:
            img = img.to(self.compute_device, non_blocking=True)
        self.images[index] = img
        return img

    def __len__(self):
        return len(self.images)


def load_video_frames(
    video_path,
    image_size,
    offload_video_to_cpu,
    img_mean=(0.5, 0.5, 0.5),
    img_std=(0.5, 0.5, 0.5),
    async_loading_frames=False,
    compute_device=torch.device("cuda"),
):
    """
    Load the video frames from video_path. The frames are resized to image_size as in
    the model and are loaded to GPU if offload_video_to_cpu=False. This is used by the demo.
    """
    is_bytes = isinstance(video_path, bytes)
    is_str = isinstance(video_path, str)
    is_mp4_path = is_str and os.path.splitext(video_path)[-1] in [".mp4", ".MP4"]
    if is_bytes or is_mp4_path:
        return load_video_frames_from_video_file(
            video_path=video_path,
            image_size=image_size,
            offload_video_to_cpu=offload_video_to_cpu,
            img_mean=img_mean,
            img_std=img_std,
            compute_device=compute_device,
        )
    elif is_str and os.path.isdir(video_path):
        return load_video_frames_from_jpg_images(
            video_path=video_path,
            image_size=image_size,
            offload_video_to_cpu=offload_video_to_cpu,
            img_mean=img_mean,
            img_std=img_std,
            async_loading_frames=async_loading_frames,
            compute_device=compute_device,
        )
    else:
        raise NotImplementedError(
            "Only MP4 video and JPEG folder are supported at this moment"
        )


def load_video_frames_from_jpg_images(
    video_path,
    image_size,
    offload_video_to_cpu,
    img_mean=(0.5, 0.5, 0.5),
    img_std=(0.5, 0.5, 0.5),
    async_loading_frames=False,
    compute_device=torch.device("cuda"),
):
    """
    Load the video frames from a directory of JPEG files ("<frame_index>.jpg" format).

    The frames are resized to image_size x image_size and are loaded to GPU if
    `offload_video_to_cpu` is `False` and to CPU if `offload_video_to_cpu` is `True`.

    You can load a frame asynchronously by setting `async_loading_frames` to `True`.
    """
    if isinstance(video_path, str) and os.path.isdir(video_path):
        jpg_folder = video_path
    else:
        raise NotImplementedError(
            "Only JPEG frames are supported at this moment. For video files, you may use "
            "ffmpeg (https://ffmpeg.org/) to extract frames into a folder of JPEG files, such as \n"
            "```\n"
            "ffmpeg -i <your_video>.mp4 -q:v 2 -start_number 0 <output_dir>/'%05d.jpg'\n"
            "```\n"
            "where `-q:v` generates high-quality JPEG frames and `-start_number 0` asks "
            "ffmpeg to start the JPEG file from 00000.jpg."
        )

    frame_names = [
        p
        for p in os.listdir(jpg_folder)
        if os.path.splitext(p)[-1] in [".jpg", ".jpeg", ".JPG", ".JPEG"]
    ]
    frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))
    num_frames = len(frame_names)
    if num_frames == 0:
        raise RuntimeError(f"no images found in {jpg_folder}")
    img_paths = [os.path.join(jpg_folder, frame_name) for frame_name in frame_names]
    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]

    if async_loading_frames:
        lazy_images = AsyncVideoFrameLoader(
            img_paths,
            image_size,
            offload_video_to_cpu,
            img_mean,
            img_std,
            compute_device,
        )
        return lazy_images, lazy_images.video_height, lazy_images.video_width

    images = torch.zeros(num_frames, 3, image_size, image_size, dtype=torch.float32)
    for n, img_path in enumerate(tqdm(img_paths, desc="frame loading (JPEG)")):
        images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
    if not offload_video_to_cpu:
        images = images.to(compute_device)
        img_mean = img_mean.to(compute_device)
        img_std = img_std.to(compute_device)
    # normalize by mean and std
    images -= img_mean
    images /= img_std
    return images, video_height, video_width


def load_video_frames_from_video_file(
    video_path,
    image_size,
    offload_video_to_cpu,
    img_mean=(0.5, 0.5, 0.5),
    img_std=(0.5, 0.5, 0.5),
    compute_device=torch.device("cuda"),
):
    """Load the video frames from a video file."""
    import decord

    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
    # Get the original video height and width
    decord.bridge.set_bridge("torch")
    video_height, video_width, _ = decord.VideoReader(video_path).next().shape
    # Iterate over all frames in the video
    images = []
    for frame in decord.VideoReader(video_path, width=image_size, height=image_size):
        images.append(frame.permute(2, 0, 1))

    images = torch.stack(images, dim=0).float() / 255.0
    if not offload_video_to_cpu:
        images = images.to(compute_device)
        img_mean = img_mean.to(compute_device)
        img_std = img_std.to(compute_device)
    # normalize by mean and std
    images -= img_mean
    images /= img_std
    return images, video_height, video_width