trajectory_construction/plan_mask.py

import os
import sys
import gc
import shutil
from pathlib import Path

# 先杀掉可能冲突的进程（可选）
os.system("pkill -f monitor_3.py")

# 获取当前脚本所在目录，并添加 Grounded_SAM2 到路径
current_dir = os.path.dirname(os.path.abspath(__file__))
grounded_sam2_path = os.path.join(current_dir, "Grounded_SAM2")
if grounded_sam2_path not in sys.path:
    sys.path.insert(0, grounded_sam2_path)

import cv2
import numpy as np
import supervision as sv
import torch
from PIL import Image
from transformers import AutoModelForZeroShotObjectDetection, AutoProcessor
from Grounded_SAM2.sam2.build_sam import build_sam2, build_sam2_video_predictor
from Grounded_SAM2.sam2.sam2_image_predictor import SAM2ImagePredictor
from Grounded_SAM2.utils.track_utils import sample_points_from_masks
import argparse

# ========================
# 配置参数
# ========================
JSON_PATH = "/mnt/prev_nas/qhy/datasets/miradata9k_trackable_objects.json"
VIDEO_ROOT_DIR = "/mnt/prev_nas/qhy/datasets/MiraData9K_download/zip/MiraData9K"
OUTPUT_DIR = "/ossfs/workspace"
PROGRESS_FILE = os.path.join(OUTPUT_DIR, "progress.json")  # ← 进度文件路径

MODEL_CFG = "sam2_hiera_l.yaml"
CHECKPOINT = "/mnt/prev_nas/qhy/MagicMotion/trajectory_construction/Grounded_SAM2/checkpoints/sam2_hiera_large.pt"
GROUNDING_MODEL_ID = "/mnt/prev_nas/qhy/MagicMotion/trajectory_construction/Grounded_SAM2/checkpoints/grounding-dino-tiny"

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
FPS = 16


def load_progress():
    """加载 progress.json，如果不存在则返回空字典"""
    if os.path.exists(PROGRESS_FILE):
        with open(PROGRESS_FILE, 'r', encoding='utf-8') as f:
            try:
                return json.load(f)
            except json.JSONDecodeError:
                print(f"⚠️ {PROGRESS_FILE} 格式错误，将重建")
                return {}
    return {}


def save_progress(progress):
    """保存进度到 JSON 文件"""
    temp_file = PROGRESS_FILE + ".tmp"
    with open(temp_file, 'w', encoding='utf-8') as f:
        json.dump(progress, f, indent=2, ensure_ascii=False)
    # 原子性替换，防止写入一半中断
    shutil.move(temp_file, PROGRESS_FILE)


def segment_core(
    text,
    video_dir,
    inference_state,
    video_predictor,
    image_predictor,
    grounding_model,
    processor,
    device="cuda"
):
    """
    核心分割逻辑：使用已初始化的模型和状态进行推理
    """
    # 获取帧名并排序
    frame_names = [
        p for p in os.listdir(video_dir)
        if os.path.splitext(p)[-1].lower() in [".jpg", ".jpeg"]
    ]
    frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))

    ann_frame_idx = 0  # 用于 prompt 的第一帧
    img_path = os.path.join(video_dir, frame_names[ann_frame_idx])
    image = Image.open(img_path).convert("RGB")

    # Grounding DINO 推理
    inputs = processor(images=image, text=text, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = grounding_model(**inputs)
    results = processor.post_process_grounded_object_detection(
        outputs,
        inputs.input_ids,
        box_threshold=0.25,
        text_threshold=0.3,
        target_sizes=[image.size[::-1]],
    )

    input_boxes = results[0]["boxes"].cpu().numpy()
    OBJECTS = results[0]["labels"]

    if len(input_boxes) == 0:
        print(f"⚠️  Grounding DINO 未检测到任何对象: '{text}'")
        return []

    # 设置图像预测器
    image_predictor.set_image(np.array(image))

    masks, _, _ = image_predictor.predict(
        point_coords=None,
        point_labels=None,
        box=input_boxes,
        multimask_output=False,
    )

    # 调整 mask 形状 (n, H, W)
    if masks.ndim == 4:
        masks = masks.squeeze(1)  # (1, 1, H, W) -> (H, W)
    elif masks.ndim == 3 and masks.shape[0] != len(OBJECTS):
        masks = masks[None]  # (H, W) -> (1, H, W)

    # 添加每个对象到视频预测器
    PROMPT_TYPE_FOR_VIDEO = "box"

    if PROMPT_TYPE_FOR_VIDEO == "point":
        all_sample_points = sample_points_from_masks(masks=masks, num_points=10)
        for obj_id, (label, points) in enumerate(zip(OBJECTS, all_sample_points), start=1):
            labels = np.ones((points.shape[0]), dtype=np.int32)
            video_predictor.add_new_points_or_box(
                inference_state=inference_state,
                frame_idx=ann_frame_idx,
                obj_id=obj_id,
                points=points,
                labels=labels,
            )
    elif PROMPT_TYPE_FOR_VIDEO == "box":
        for obj_id, (label, box) in enumerate(zip(OBJECTS, input_boxes), start=1):
            video_predictor.add_new_points_or_box(
                inference_state=inference_state,
                frame_idx=ann_frame_idx,
                obj_id=obj_id,
                box=box,
            )
    elif PROMPT_TYPE_FOR_VIDEO == "mask":
        for obj_id, (label, mask) in enumerate(zip(OBJECTS, masks), start=1):
            video_predictor.add_new_mask(
                inference_state=inference_state,
                frame_idx=ann_frame_idx,
                obj_id=obj_id,
                mask=mask,
            )
    else:
        raise NotImplementedError("Only support 'point', 'box', or 'mask'.")

    # 传播所有帧
    video_segments = {}
    for out_frame_idx, out_obj_ids, out_mask_logits in video_predictor.propagate_in_video(inference_state):
        video_segments[out_frame_idx] = {
            out_obj_id: (out_mask_logits[i] > 0.0).cpu().numpy()[0]
            for i, out_obj_id in enumerate(out_obj_ids)
        }

    # 可视化结果
    annotated_frames = []
    for frame_idx in range(len(frame_names)):
        frame_path = os.path.join(video_dir, frame_names[frame_idx])
        if not os.path.exists(frame_path):
            continue
        img = cv2.imread(frame_path)
        if img is None:
            continue

        if frame_idx not in video_segments or len(video_segments[frame_idx]) == 0:
            annotated_frames.append(np.zeros_like(img))
            continue

        segments = video_segments[frame_idx]
        object_ids = list(segments.keys())
        masks_list = list(segments.values())
        masks_array = np.stack(masks_list, axis=0)  # (n, h, w)

        detections = sv.Detections(
            xyxy=sv.mask_to_xyxy(masks_array),
            mask=masks_array,
            class_id=np.array(object_ids, dtype=int),
        )
        mask_annotator = sv.MaskAnnotator()
        annotated_frame = mask_annotator.annotate(
            scene=np.zeros_like(img), detections=detections
        )
        annotated_frames.append(annotated_frame)

    return annotated_frames


def save_video(frames, output_path, fps=16):
    """保存帧为 MP4 视频"""
    if not frames:
        print(f"⚠️ 无帧可保存: {output_path}")
        return
    height, width, _ = frames[0].shape
    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    os.makedirs(os.path.dirname(output_path), exist_ok=True)
    video_writer = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
    for frame in frames:
        video_writer.write(frame)
    video_writer.release()


def extract_all_frames(video_path, output_dir=None, image_format="jpg"):
    """
    提取视频所有帧
    """
    video_path = str(video_path)
    if not os.path.exists(video_path):
        raise FileNotFoundError(f"视频不存在: {video_path}")

    if output_dir is None:
        video_stem = Path(video_path).stem
        parent_dir = Path(video_path).parent
        output_dir = str(parent_dir / f"{video_stem}_frames")

    if os.path.exists(output_dir):
        shutil.rmtree(output_dir)
    os.makedirs(output_dir, exist_ok=False)

    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise IOError(f"无法打开视频: {video_path}")

    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    print(f"📽️ 处理视频: {video_path} ({total_frames} 帧)")

    saved_count = 0
    frame_idx = 0
    while True:
        ret, frame = cap.read()
        if not ret:
            break
        save_path = os.path.join(output_dir, f"{frame_idx:06d}.{image_format}")
        success = cv2.imwrite(save_path, frame)
        if success:
            saved_count += 1
        frame_idx += 1
    cap.release()

    print(f"✅ 已提取 {saved_count} 帧到: {output_dir}")
    return output_dir


def main():
    # 加载进度
    progress = load_progress()
    print(f"📁 已加载 {len(progress)} 条处理记录")

    # 加载任务列表
    with open(JSON_PATH, 'r', encoding='utf-8') as f:
        data = json.load(f)
    print(f"📋 共 {len(data)} 个视频待处理")
    # ========================
    # 初始化模型（只加载一次）
    # ========================
    print("🚀 正在加载模型...")
    torch.autocast(device_type="cuda", enabled=(DEVICE=="cuda"), dtype=torch.bfloat16).__enter__()

    if torch.cuda.is_available():
        torch.backends.cuda.matmul.allow_tf32 = True
        torch.backends.cudnn.allow_tf32 = True

    # 构建视频和图像预测器
    video_predictor = build_sam2_video_predictor(MODEL_CFG, CHECKPOINT).to(DEVICE)
    sam2_image_model = build_sam2(MODEL_CFG, CHECKPOINT).to(DEVICE)
    image_predictor = SAM2ImagePredictor(sam2_image_model)

    # 加载 Grounding DINO
    processor = AutoProcessor.from_pretrained(GROUNDING_MODEL_ID)
    grounding_model = AutoModelForZeroShotObjectDetection.from_pretrained(GROUNDING_MODEL_ID).to(DEVICE)

    print("✅ 模型加载完成")

    # ========================
    # 加载数据并处理
    # ========================
    with open(JSON_PATH, 'r', encoding='utf-8') as f:
        data = json.load(f)

    os.makedirs(OUTPUT_DIR, exist_ok=True)
    print(f"Loaded {len(data)} videos")

    for item in data:
        video_path = item["video_path"]          # e.g., MiraData9K/xxx/yyy.mp4
        trackable_objects = item["trackable_objects"]
        output_video_path = os.path.join(OUTPUT_DIR, video_path.replace(".mp4", ".mp4"))
        output_video_path = output_video_path.replace("MiraData9K/", "")  # 清理路径
        local_video_path = os.path.join(VIDEO_ROOT_DIR, video_path)

        # 输出文件存在且未在 progress 中？→ 补录为 done
        if os.path.exists(output_video_path) and video_path not in progress:
            progress[video_path] = {"status": "done"}
            print(f"🔍 发现已存在结果，标记为完成: {video_path}")
            continue

        # 检查是否已完成
        if progress.get(video_path, {}).get("status") == "done":
            print(f"⏭️ 已完成，跳过: {video_path}")
            continue
        
        if not trackable_objects:
            progress[video_path] = {"status": "skipped"}
            print(f"⏭️ 无跟踪对象，跳过: {video_path}")
            continue
        

        print(f"🎬 Processing: {video_path}")
        print(f"   Objects: {trackable_objects}")

        try:
            # 1. 提取帧
            frame_dir = extract_all_frames(local_video_path)

            # 2. 初始化视频状态（关键：每次新建）
            inference_state = video_predictor.init_state(video_path=frame_dir)

            # 3. 执行分割
            text_prompt = ".".join(trackable_objects) + "."
            annotated_frames = segment_core(
                text=text_prompt,
                video_dir=frame_dir,
                inference_state=inference_state,
                video_predictor=video_predictor,
                image_predictor=image_predictor,
                grounding_model=grounding_model,
                processor=processor,
                device=DEVICE
            )

            if not annotated_frames:
                print(f"❌ 分割结果为空: {video_path}")
                continue

            # 4. 保存轨迹视频
            save_video(annotated_frames, output_video_path, fps=FPS)
            progress[video_path] = {"status": "done"}
            print(f"✅ 成功保存: {output_video_path}")

        except Exception as e:
            print(f"❌ 处理失败 {video_path}: {str(e)}")
            progress[video_path] = {"status": "failed", "error": str(e)}

        finally:
            # ✅ 关键：释放 inference_state 和临时资源
            if 'inference_state' in locals():
                video_predictor.reset_state(inference_state)
                del inference_state
            if 'frame_dir' in locals() and os.path.exists(frame_dir):
                shutil.rmtree(frame_dir)
            torch.cuda.empty_cache()
            gc.collect()
        save_progress(progress)

    print("🎉 所有视频处理完成！")


if __name__ == "__main__":
    import json
    main()