Upload 3 files

play_video.py

@@ -0,0 +1,158 @@
+import os
+import glob
+import subprocess
+import gradio as gr
+
+CI_VID_DIR = "ci_vid_SC_test"
+CACHE_DIR = "play_video/merged_previews"
+
+FFMPEG_PATH = r"D:\ffmpeg-8.0.1-essentials_build\bin\ffmpeg.exe"
+
+os.makedirs(CACHE_DIR, exist_ok=True)
+
+def scan_data():
+    data_map = {}
+    if not os.path.exists(CI_VID_DIR): return data_map
+    for cp in sorted(glob.glob(os.path.join(CI_VID_DIR, "chunk_*"))):
+        chunk = os.path.basename(cp)
+        data_map[chunk] = {}
+        for vp in sorted(glob.glob(os.path.join(cp, "*"))):
+            if not os.path.isdir(vp): continue
+            vid = os.path.basename(vp)
+            data_map[chunk][vid] = {}
+            for cmp in sorted(glob.glob(os.path.join(vp, "*"))):
+                if not os.path.isdir(cmp): continue
+                cam = os.path.basename(cmp)
+                shots = sorted(
+                    [os.path.basename(f).replace('.mp4', '') for f in glob.glob(os.path.join(cmp, "*.mp4")) if os.path.basename(f).replace('.mp4', '').isdigit()],
+                    key=lambda x: int(x)
+                )
+                if shots:
+                    data_map[chunk][vid][cam] = shots
+    return data_map
+
+DATA_MAP = scan_data()
+
+def get_init():
+    c = list(DATA_MAP.keys())[0] if DATA_MAP else "-"
+    v = list(DATA_MAP.get(c, {}).keys())[0] if DATA_MAP.get(c) else "-"
+    cam = list(DATA_MAP.get(c, {}).get(v, {}).keys())[0] if DATA_MAP.get(c, {}).get(v) else "-"
+    return list(DATA_MAP.keys()), c, list(DATA_MAP.get(c, {}).keys()), v, list(DATA_MAP.get(c, {}).get(v, {}).keys()), cam
+
+CHUNKS, C, VIDS, V, CAMS, CAM = get_init()
+
+def get_choices_and_val(lst):
+    return gr.update(choices=lst, value=lst[0] if lst else "-")
+
+# 【新增】确保视频为H.264格式并缓存
+def ensure_h264_cache(orig_path, chunk, vid, cam, shot_name):
+    if not orig_path: return None
+    
+    # 构造缓存路径
+    out_name = f"h264_orig_{chunk}_{vid}_{cam}_{shot_name}.mp4"
+    out_path = os.path.join(CACHE_DIR, out_name)
+    
+    # 如果已经转码过了，直接返回，秒开
+    if os.path.exists(out_path):
+        return out_path
+        
+    # 如果没转过，用 FFmpeg 快速转码为 H.264
+    cmd = [
+        FFMPEG_PATH, '-y', '-i', orig_path, 
+        '-c:v', 'libx264', '-preset', 'ultrafast', '-crf', '28', 
+        '-c:a', 'copy', out_path
+    ]
+    subprocess.run(cmd, capture_output=True)
+    
+    return out_path if os.path.exists(out_path) else None
+
+def update_4_shots(chunk, vid, cam):
+    if "-" in [chunk, vid, cam]: 
+        return [None, None, None, None]
+    
+    search_path = os.path.join(CI_VID_DIR, chunk, vid, cam, "*.mp4")
+    shots = sorted(
+        [p for p in glob.glob(search_path) if os.path.basename(p).replace('.mp4','').isdigit()],
+        key=lambda x: int(os.path.basename(x).split('.')[0])
+    )
+    
+    out = []
+    for i in range(4):
+        if i < len(shots):
+            # 获取原始视频路径，并通过转码函数拿到 H.264 版本的缓存路径
+            orig_path = shots[i]
+            shot_name = os.path.basename(orig_path).replace('.mp4', '')
+            safe_path = ensure_h264_cache(orig_path, chunk, vid, cam, shot_name)
+            out.append(safe_path)
+        else:
+            out.append(None)
+    return out
+
+def chunk_update(chunk):
+    vids = list(DATA_MAP.get(chunk, {}).keys())
+    v_def = vids[0] if vids else "-"
+    cams = list(DATA_MAP.get(chunk, {}).get(v_def, {}).keys())
+    c_def = cams[0] if cams else "-"
+    
+    s1, s2, s3, s4 = update_4_shots(chunk, v_def, c_def)
+    return get_choices_and_val(vids), get_choices_and_val(cams), s1, s2, s3, s4
+
+def vid_update(chunk, vid):
+    cams = list(DATA_MAP.get(chunk, {}).get(vid, {}).keys())
+    c_def = cams[0] if cams else "-"
+    
+    s1, s2, s3, s4 = update_4_shots(chunk, vid, c_def)
+    return get_choices_and_val(cams), s1, s2, s3, s4
+
+def cam_update(chunk, vid, cam):
+    s1, s2, s3, s4 = update_4_shots(chunk, vid, cam)
+    return s1, s2, s3, s4
+
+def concat_original_shots(chunk, vid, cam):
+    if "-" in [chunk, vid, cam]: return None
+    paths = update_4_shots(chunk, vid, cam)
+    inputs = [p for p in paths if p is not None]
+    if not inputs: return None
+    
+    out_path = os.path.join(CACHE_DIR, f"concat_orig_{chunk}_{vid}_{cam}.mp4")
+    if os.path.exists(out_path):
+        return out_path
+
+    flt = "".join([f"[{i}:v]scale=1280:720:force_original_aspect_ratio=decrease,pad=1280:720:(ow-iw)/2:(oh-ih)/2,setsar=1,fps=24[v{i}];" for i in range(len(inputs))])
+    concat = "".join([f"[v{i}]" for i in range(len(inputs))]) + f"concat=n={len(inputs)}:v=1:a=0[outv]"
+    cmd = [FFMPEG_PATH, "-y", *sum([["-i", inp] for inp in inputs], []), "-filter_complex", flt + concat, "-map", "[outv]", "-c:v", "libx264", "-preset", "ultrafast", out_path]
+    subprocess.run(cmd, check=True, capture_output=True)
+    
+    return out_path
+
+with gr.Blocks() as app:
+    gr.Markdown("# 🎬 原始序列视频拼接与查看器")
+    
+    with gr.Row():
+        chunk_dd = gr.Dropdown(choices=CHUNKS, label="1. Chunk", value=C)
+        vid_dd   = gr.Dropdown(choices=VIDS, label="2. Video ID", value=V)
+        cam_dd   = gr.Dropdown(choices=CAMS, label="3. Camera ID", value=CAM)
+    
+    gr.Markdown("---")
+    gr.Markdown("### 🎥 原始 Shot 预览 (按顺序自动加载，最多展示4个)")
+    with gr.Row():
+        shot_vids = [gr.Video(label=f"Shot {i+1}") for i in range(4)]
+    
+    gr.Markdown("---")
+    with gr.Row():
+        btn_concat_orig = gr.Button("🔗 拼接以上原始视频", variant="primary")
+        v_orig_concat = gr.Video(label="拼接结果")
+        
+    chunk_dd.change(chunk_update, [chunk_dd], [vid_dd, cam_dd] + shot_vids)
+    vid_dd.change(vid_update, [chunk_dd, vid_dd], [cam_dd] + shot_vids)
+    cam_dd.change(cam_update, [chunk_dd, vid_dd, cam_dd], shot_vids)
+    
+    btn_concat_orig.click(concat_original_shots, [chunk_dd, vid_dd, cam_dd], [v_orig_concat])
+    
+    app.load(lambda c, v, cam: tuple(update_4_shots(c, v, cam)), [chunk_dd, vid_dd, cam_dd], shot_vids)
+
+if __name__ == "__main__":
+    print(f"Link: http://127.0.0.1:7860/")
+    # 放开限制
+    allowed = [os.path.abspath(CI_VID_DIR).replace("\\", "/"), os.path.abspath(CACHE_DIR).replace("\\", "/"), "/workspace", os.path.abspath(".")]
+    app.launch(server_name="0.0.0.0", server_port=7860, share=True, allowed_paths=allowed)

quality_check.py

@@ -0,0 +1,354 @@
+import os
+import glob
+import subprocess
+import gradio as gr
+import cv2
+import numpy as np
+import json
+
+STAGE3_DIR = "test_removal_0_39/stage3_prop_masks"
+STAGE4_DIR = "test_removal_0_39/stage4_inpainted"
+CI_VID_DIR = "ci_vid_SC_test"
+CACHE_DIR = "outputs/merged_previews"
+ANNOTATION_FILE = "removal_annotations.json"
+
+FFMPEG_PATH = r"D:\ffmpeg-8.0.1-essentials_build\bin\ffmpeg.exe"
+
+os.makedirs(CACHE_DIR, exist_ok=True)
+
+# ======================= 【层级化 JSON 读写逻辑】 =======================
+def load_annotations():
+    """读取层级化 JSON 标注记录"""
+    if os.path.exists(ANNOTATION_FILE):
+        try:
+            with open(ANNOTATION_FILE, mode='r', encoding='utf-8') as f:
+                return json.load(f)
+        except Exception:
+            return {}
+    return {}
+
+def save_annotation(chunk, vid, cam, shot, obj, rating):
+    """保存标注到层级化 JSON 文件"""
+    if not rating:
+        return "⚠️ Please select a rating first!"
+    if "-" in [chunk, vid, cam, shot, obj]:
+        return "⚠️ Please select a valid clip!"
+
+    anns = load_annotations()
+    
+    if chunk not in anns: anns[chunk] = {}
+    if vid not in anns[chunk]: anns[chunk][vid] = {}
+    if cam not in anns[chunk][vid]: anns[chunk][vid][cam] = {}
+    if shot not in anns[chunk][vid][cam]: anns[chunk][vid][cam][shot] = {}
+    
+    anns[chunk][vid][cam][shot][obj] = rating 
+
+    with open(ANNOTATION_FILE, mode='w', encoding='utf-8') as f:
+        json.dump(anns, f, ensure_ascii=False, indent=4)
+
+    return f"✅ Saved: `{obj}` marked as **{rating}**"
+
+def sync_annotation_ui(chunk, vid, cam, shot, obj):
+    """同步 UI 状态 (安全地从嵌套字典中读取)"""
+    anns = load_annotations()
+    current_rating = anns.get(chunk, {}).get(vid, {}).get(cam, {}).get(shot, {}).get(obj, None)
+    return gr.update(value=current_rating), ""
+# =====================================================================
+
+
+def scan_data():
+    data_map = {}
+    if not os.path.exists(STAGE4_DIR): return data_map
+    for cp in sorted(glob.glob(os.path.join(STAGE4_DIR, "chunk_*"))):
+        chunk = os.path.basename(cp)
+        data_map[chunk] = {}
+        for vp in sorted(glob.glob(os.path.join(cp, "*"))):
+            if not os.path.isdir(vp): continue
+            vid = os.path.basename(vp)
+            data_map[chunk][vid] = {}
+            for cmp in sorted(glob.glob(os.path.join(vp, "*"))):
+                if not os.path.isdir(cmp): continue
+                cam = os.path.basename(cmp)
+                data_map[chunk][vid][cam] = {}
+                for sp in sorted(glob.glob(os.path.join(cmp, "*"))):
+                    if not os.path.isdir(sp): continue
+                    shot = os.path.basename(sp)
+                    if not shot.isdigit(): continue
+                    objs = [os.path.basename(o).replace('.mp4', '') for o in glob.glob(os.path.join(sp, "*.mp4"))]
+                    if objs: data_map[chunk][vid][cam][shot] = sorted(objs)
+    return data_map
+
+ORIGINAL_DATA_MAP = scan_data()
+
+def get_init(data_map):
+    c = list(data_map.keys())[0] if data_map else "-"
+    v = list(data_map.get(c, {}).keys())[0] if data_map.get(c) else "-"
+    cam = list(data_map.get(c, {}).get(v, {}).keys())[0] if data_map.get(c, {}).get(v) else "-"
+    s = list(data_map.get(c, {}).get(v, {}).get(cam, {}).keys())[0] if data_map.get(c, {}).get(v, {}).get(cam) else "-"
+    o = data_map.get(c, {}).get(v, {}).get(cam, {}).get(s, [])[0] if data_map.get(c, {}).get(v, {}).get(cam, {}).get(s) else "-"
+    return list(data_map.keys()), c, list(data_map.get(c, {}).keys()), v, list(data_map.get(c, {}).get(v, {}).keys()), cam, list(data_map.get(c, {}).get(v, {}).get(cam, {}).keys()), s, data_map.get(c, {}).get(v, {}).get(cam, {}).get(s, []), o
+
+
+# ======================= 【新增】动态计数与刷新逻辑 =======================
+def get_filter_choices_and_val(base_filter="All"):
+    """计算各标签的数量，并生成带数字的 Radio 选项"""
+    anns = load_annotations()
+    counts = {"All": 0, "Unlabeled": 0, "High": 0, "Medium": 0, "Low": 0}
+    
+    for c, v_dict in ORIGINAL_DATA_MAP.items():
+        for v, cam_dict in v_dict.items():
+            for cam, s_dict in cam_dict.items():
+                for s, o_list in s_dict.items():
+                    for o in o_list:
+                        counts["All"] += 1
+                        val = anns.get(c, {}).get(v, {}).get(cam, {}).get(s, {}).get(o, "")
+                        if not val:
+                            counts["Unlabeled"] += 1
+                        elif val in counts:
+                            counts[val] += 1
+                            
+    choices = [
+        f"All ({counts['All']})",
+        f"Unlabeled ({counts['Unlabeled']})",
+        f"High ({counts['High']})",
+        f"Medium ({counts['Medium']})",
+        f"Low ({counts['Low']})"
+    ]
+    
+    # 找到当前应该选中的那个项 (模糊匹配前缀)
+    current_val = choices[0]
+    for choice in choices:
+        if choice.startswith(base_filter):
+            current_val = choice
+            break
+            
+    return choices, current_val
+
+def apply_filter(base_filter):
+    """基于纯净标签 (如 'All') 进行过滤"""
+    if base_filter == "All":
+        return ORIGINAL_DATA_MAP
+    
+    anns = load_annotations()
+    filtered = {}
+    for c, v_dict in ORIGINAL_DATA_MAP.items():
+        for v, cam_dict in v_dict.items():
+            for cam, s_dict in cam_dict.items():
+                for s, o_list in s_dict.items():
+                    valid_objs = []
+                    for o in o_list:
+                        val = anns.get(c, {}).get(v, {}).get(cam, {}).get(s, {}).get(o, "")
+                        if base_filter == "Unlabeled" and not val:
+                            valid_objs.append(o)
+                        elif val == base_filter:
+                            valid_objs.append(o)
+                            
+                    if valid_objs:
+                        if c not in filtered: filtered[c] = {}
+                        if v not in filtered[c]: filtered[c][v] = {}
+                        if cam not in filtered[c][v]: filtered[c][v][cam] = {}
+                        filtered[c][v][cam][s] = valid_objs
+    return filtered
+
+def refresh_ui(current_filter_label):
+    """统一的刷新事件入口：重新统计数字 -> 重新过滤数据树 -> 刷新所有 UI"""
+    # 提取纯净标签 (把 "Unlabeled (42)" 截取成 "Unlabeled")
+    base_filter = "All" if not current_filter_label else current_filter_label.split(" ")[0]
+    
+    choices, new_radio_val = get_filter_choices_and_val(base_filter)
+    filtered_map = apply_filter(base_filter)
+    
+    _, c, _, v, _, cam, _, s, _, o = get_init(filtered_map)
+    
+    c_list = list(filtered_map.keys())
+    v_list = list(filtered_map.get(c, {}).keys()) if c != "-" else []
+    cam_list = list(filtered_map.get(c, {}).get(v, {}).keys()) if v != "-" else []
+    s_list = list(filtered_map.get(c, {}).get(v, {}).get(cam, {}).keys()) if cam != "-" else []
+    o_list = filtered_map.get(c, {}).get(v, {}).get(cam, {}).get(s, []) if s != "-" else []
+    vr = get_valid_replace_shots(filtered_map, c, v, cam, o)
+    
+    return (
+        filtered_map,
+        gr.update(choices=choices, value=new_radio_val), # 更新带有最新数字的选项
+        get_choices_and_val(c_list),
+        get_choices_and_val(v_list),
+        get_choices_and_val(cam_list),
+        get_choices_and_val(s_list),
+        get_choices_and_val(o_list),
+        gr.update(choices=vr, value=vr if vr else "-", label=f"Use replacement clips for [{o if o!='-' else 'None'}]")
+    )
+# =====================================================================
+
+def get_choices_and_val(lst):
+    return gr.update(choices=lst, value=lst[0] if lst else "-")
+
+def get_valid_replace_shots(data_map, chunk, vid, cam, obj):
+    if "-" in [chunk, vid, cam, obj]: return []
+    return [s for s in list(data_map.get(chunk, {}).get(vid, {}).get(cam, {}).keys()) if obj in data_map.get(chunk, {}).get(vid, {}).get(cam, {}).get(s, [])]
+
+def chunk_update(state_map, chunk):
+    vids = list(state_map.get(chunk, {}).keys())
+    cams = list(state_map.get(chunk, {}).get(vids[0] if vids else "-", {}).keys())
+    shots = list(state_map.get(chunk, {}).get(vids[0] if vids else "-", {}).get(cams[0] if cams else "-", {}).keys())
+    objs = state_map.get(chunk, {}).get(vids[0] if vids else "-", {}).get(cams[0] if cams else "-", {}).get(shots[0] if shots else "-", [])
+    vr = get_valid_replace_shots(state_map, chunk, vids[0] if vids else "-", cams[0] if cams else "-", objs[0] if objs else "-")
+    return get_choices_and_val(vids), get_choices_and_val(cams), get_choices_and_val(shots), get_choices_and_val(objs), gr.update(choices=vr, value=vr, label=f"Use replacement clips for [{objs[0] if objs else '-'}]")
+
+def vid_update(state_map, chunk, vid):
+    cams = list(state_map.get(chunk, {}).get(vid, {}).keys())
+    shots = list(state_map.get(chunk, {}).get(vid, {}).get(cams[0] if cams else "-", {}).keys())
+    objs = state_map.get(chunk, {}).get(vid, {}).get(cams[0] if cams else "-", {}).get(shots[0] if shots else "-", [])
+    vr = get_valid_replace_shots(state_map, chunk, vid, cams[0] if cams else "-", objs[0] if objs else "-")
+    return get_choices_and_val(cams), get_choices_and_val(shots), get_choices_and_val(objs), gr.update(choices=vr, value=vr, label=f"Use replacement clips for [{objs[0] if objs else '-'}]")
+
+def cam_update(state_map, chunk, vid, cam):
+    shots = list(state_map.get(chunk, {}).get(vid, {}).get(cam, {}).keys())
+    objs = state_map.get(chunk, {}).get(vid, {}).get(cam, {}).get(shots[0] if shots else "-", [])
+    vr = get_valid_replace_shots(state_map, chunk, vid, cam, objs[0] if objs else "-")
+    return get_choices_and_val(shots), get_choices_and_val(objs), gr.update(choices=vr, value=vr, label=f"Use replacement clips for [{objs[0] if objs else '-'}]")
+
+def shot_update(state_map, chunk, vid, cam, shot):
+    objs = state_map.get(chunk, {}).get(vid, {}).get(cam, {}).get(shot, [])
+    vr = get_valid_replace_shots(state_map, chunk, vid, cam, objs[0] if objs else "-")
+    return get_choices_and_val(objs), gr.update(choices=vr, value=vr, label=f"Use replacement clips for [{objs[0] if objs else '-'}]")
+
+def obj_update(state_map, chunk, vid, cam, obj):
+    vr = get_valid_replace_shots(state_map, chunk, vid, cam, obj)
+    return gr.update(choices=vr, value=vr, label=f"Use replacement clips for [{obj}]")
+
+def create_overlay_video(orig_path, mask_dir, out_path):
+    if os.path.exists(out_path): return out_path
+    
+    cap = cv2.VideoCapture(orig_path)
+    fps = cap.get(cv2.CAP_PROP_FPS) or 24
+    orig_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    orig_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    
+    scale = 0.5
+    w = int((orig_w * scale) // 2 * 2)
+    h = int((orig_h * scale) // 2 * 2)
+
+    mask_files = sorted(glob.glob(os.path.join(mask_dir, "*.png")))
+    temp_path = out_path.replace('.mp4', '_temp.mp4')
+    out = cv2.VideoWriter(temp_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+    
+    idx = 0
+    while True:
+        ret, frame = cap.read()
+        if not ret: break
+        
+        frame = cv2.resize(frame, (w, h), interpolation=cv2.INTER_AREA)
+        if idx < len(mask_files):
+            mask = cv2.imread(mask_files[idx], cv2.IMREAD_GRAYSCALE)
+            if mask is not None:
+                mask = cv2.resize(mask, (w, h), interpolation=cv2.INTER_NEAREST)
+                frame[mask > 128] = frame[mask > 128] * 0.5 + np.array([0, 0, 255], dtype=np.uint8) * 0.5
+        out.write(frame)
+        idx += 1
+        
+    cap.release()
+    out.release()
+    
+    subprocess.run([FFMPEG_PATH, '-y', '-i', temp_path, '-c:v', 'libx264', '-preset', 'ultrafast', '-crf', '32', out_path], capture_output=True)
+    if os.path.exists(temp_path): os.remove(temp_path)
+    return out_path
+
+def load_preview(chunk, vid, cam, shot, obj):
+    if not all([chunk, vid, cam, shot, obj]) or "-" in [chunk, vid, cam, shot, obj]: return None, None
+    orig = os.path.join(CI_VID_DIR, chunk, vid, cam, f"{shot}.mp4")
+    edit = os.path.join(STAGE4_DIR, chunk, vid, cam, str(shot), f"{obj}.mp4")
+    mask_dir = os.path.join(STAGE3_DIR, chunk, vid, cam, str(shot), str(obj))
+    orig_to_show = create_overlay_video(orig, mask_dir, os.path.join(CACHE_DIR, f"ov_{chunk}_{vid}_{cam}_{shot}_{obj}.mp4")) if os.path.exists(orig) and os.path.exists(mask_dir) else (orig if os.path.exists(orig) else None)
+    return orig_to_show, (edit if os.path.exists(edit) else None)
+
+def merge_videos(chunk, vid, cam, obj, replace_shots):
+    if not all([chunk, vid, cam, obj]) or "-" in [chunk, vid, cam, obj]: return None
+    all_orig = sorted([p for p in glob.glob(os.path.join(CI_VID_DIR, chunk, vid, cam, "*.mp4")) if os.path.basename(p).replace('.mp4','').isdigit()], key=lambda x: int(os.path.basename(x).split('.')[0]))
+    out_path = os.path.join(CACHE_DIR, f"merged_{chunk}_{vid}_{cam}_{obj}.mp4")
+    inputs = []
+    for p in all_orig:
+        sid = os.path.basename(p).replace('.mp4','')
+        edit_p = os.path.join(STAGE4_DIR, chunk, vid, cam, sid, f"{obj}.mp4")
+        inputs.append(edit_p if (sid in replace_shots and os.path.exists(edit_p)) else p)
+    if not inputs: return None
+    flt = "".join([f"[{i}:v]scale=1280:720:force_original_aspect_ratio=decrease,pad=1280:720:(ow-iw)/2:(oh-ih)/2,setsar=1,fps=24[v{i}];" for i in range(len(inputs))])
+    concat = "".join([f"[v{i}]" for i in range(len(inputs))]) + f"concat=n={len(inputs)}:v=1:a=0[outv]"
+    cmd = [FFMPEG_PATH, "-y", *sum([["-i", inp] for inp in inputs], []), "-filter_complex", flt + concat, "-map", "[outv]", "-c:v", "libx264", "-preset", "ultrafast", out_path]
+    subprocess.run(cmd, check=True, capture_output=True)
+    return out_path
+
+# 获取初始标签和数字
+INIT_CHOICES, INIT_VAL = get_filter_choices_and_val("All")
+CHUNKS, C, VIDS, V, CAMS, CAM, SHOTS, S, OBJS, O = get_init(ORIGINAL_DATA_MAP)
+
+with gr.Blocks() as app:
+    state_map = gr.State(ORIGINAL_DATA_MAP)
+
+    gr.Markdown("# 🎬 Inconsistency Quality Check Panel")
+    
+    with gr.Row(variant="panel"):
+        filter_radio = gr.Radio(
+            choices=INIT_CHOICES, 
+            value=INIT_VAL, 
+            label="🔎 Filter Status"
+        )
+    
+    with gr.Row():
+        chunk_dd = gr.Dropdown(choices=CHUNKS, label="1. Chunk", value=C)
+        vid_dd   = gr.Dropdown(choices=VIDS, label="2. Video ID", value=V)
+        cam_dd   = gr.Dropdown(choices=CAMS, label="3. Camera ID", value=CAM)
+    gr.Markdown("---")
+    with gr.Row():
+        with gr.Column():
+            with gr.Row():
+                shot_dd = gr.Dropdown(choices=SHOTS, label="4. Shot", value=S)
+                obj_dd  = gr.Dropdown(choices=OBJS, label="5. Object", value=O)
+            
+            gr.Markdown("### ✍️ Evaluation")
+            with gr.Row(variant="panel"):
+                rating_radio = gr.Radio(choices=["High", "Medium", "Low"], label="6. Rate Removal Quality")
+                with gr.Column():
+                    save_btn = gr.Button("💾 Save Rating", variant="primary")
+                    refresh_btn = gr.Button("🔄 Refresh List & Stats", variant="secondary") # 【新增刷新按钮】
+            save_status = gr.Markdown("")
+
+            v_orig = gr.Video(label="Original w/ Mask")
+            v_edit = gr.Video(label="Inpainted Result")
+
+        with gr.Column():
+            replace = gr.CheckboxGroup(choices=get_valid_replace_shots(ORIGINAL_DATA_MAP, C, V, CAM, O), label=f"Use replacement clips for [{O}]", value=get_valid_replace_shots(ORIGINAL_DATA_MAP, C, V, CAM, O))
+            btn     = gr.Button("🚀 Generate Sequence Preview", variant="secondary")
+            v_res   = gr.Video(label="Sequence Preview")
+
+    # ================= 绑定事件 =================
+    
+    # 0. 刷新按钮 与 筛选器 共用同一套更新逻辑
+    refresh_outputs = [state_map, filter_radio, chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd, replace]
+    filter_radio.change(refresh_ui, [filter_radio], refresh_outputs)
+    refresh_btn.click(refresh_ui, [filter_radio], refresh_outputs)
+
+    # 1. 列表联级更新 (毫秒级)
+    chunk_dd.change(chunk_update, [state_map, chunk_dd], [vid_dd, cam_dd, shot_dd, obj_dd, replace])
+    vid_dd.change(vid_update, [state_map, chunk_dd, vid_dd], [cam_dd, shot_dd, obj_dd, replace])
+    cam_dd.change(cam_update, [state_map, chunk_dd, vid_dd, cam_dd], [shot_dd, obj_dd, replace])
+    shot_dd.change(shot_update, [state_map, chunk_dd, vid_dd, cam_dd, shot_dd], [obj_dd, replace])
+    obj_dd.change(obj_update, [state_map, chunk_dd, vid_dd, cam_dd, obj_dd], [replace])
+    
+    # 2. 优先刷新 UI 状态 (毫秒级) 
+    shot_dd.change(sync_annotation_ui, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [rating_radio, save_status])
+    obj_dd.change(sync_annotation_ui, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [rating_radio, save_status])
+    app.load(sync_annotation_ui, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [rating_radio, save_status])
+    
+    save_btn.click(save_annotation, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd, rating_radio], [save_status])
+    
+    # 3. 最后加载视频 (耗时)
+    shot_dd.change(load_preview, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [v_orig, v_edit])
+    obj_dd.change(load_preview, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [v_orig, v_edit])
+    app.load(load_preview, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [v_orig, v_edit])
+    
+    # 合成全序列视频
+    btn.click(merge_videos, [chunk_dd, vid_dd, cam_dd, obj_dd, replace], [v_res])
+
+if __name__ == "__main__":
+    print(f"Link: http://127.0.0.1:7860/")
+    app.launch(server_name="0.0.0.0", server_port=7860)

viewer_app_win.py

@@ -0,0 +1,189 @@
+import os
+import glob
+import subprocess
+import gradio as gr
+import cv2
+import numpy as np
+
+STAGE3_DIR = "test_removal_0_39/stage3_prop_masks"
+STAGE4_DIR = "test_removal_0_39/stage4_inpainted"
+CI_VID_DIR = "ci_vid_SC_test"
+CACHE_DIR = "outputs/merged_previews"
+
+FFMPEG_PATH = r"D:\ffmpeg-8.0.1-essentials_build\bin\ffmpeg.exe"
+
+os.makedirs(CACHE_DIR, exist_ok=True)
+
+def scan_data():
+    data_map = {}
+    if not os.path.exists(STAGE4_DIR): return data_map
+    for cp in sorted(glob.glob(os.path.join(STAGE4_DIR, "chunk_*"))):
+        chunk = os.path.basename(cp)
+        data_map[chunk] = {}
+        for vp in sorted(glob.glob(os.path.join(cp, "*"))):
+            if not os.path.isdir(vp): continue
+            vid = os.path.basename(vp)
+            data_map[chunk][vid] = {}
+            for cmp in sorted(glob.glob(os.path.join(vp, "*"))):
+                if not os.path.isdir(cmp): continue
+                cam = os.path.basename(cmp)
+                data_map[chunk][vid][cam] = {}
+                for sp in sorted(glob.glob(os.path.join(cmp, "*"))):
+                    if not os.path.isdir(sp): continue
+                    shot = os.path.basename(sp)
+                    if not shot.isdigit(): continue
+                    objs = [os.path.basename(o).replace('.mp4', '') for o in glob.glob(os.path.join(sp, "*.mp4"))]
+                    if objs: data_map[chunk][vid][cam][shot] = sorted(objs)
+    return data_map
+
+DATA_MAP = scan_data()
+
+def get_init():
+    c = list(DATA_MAP.keys())[0] if DATA_MAP else "-"
+    v = list(DATA_MAP.get(c, {}).keys())[0] if DATA_MAP.get(c) else "-"
+    cam = list(DATA_MAP.get(c, {}).get(v, {}).keys())[0] if DATA_MAP.get(c, {}).get(v) else "-"
+    s = list(DATA_MAP.get(c, {}).get(v, {}).get(cam, {}).keys())[0] if DATA_MAP.get(c, {}).get(v, {}).get(cam) else "-"
+    o = DATA_MAP.get(c, {}).get(v, {}).get(cam, {}).get(s, [])[0] if DATA_MAP.get(c, {}).get(v, {}).get(cam, {}).get(s) else "-"
+    return list(DATA_MAP.keys()), c, list(DATA_MAP.get(c, {}).keys()), v, list(DATA_MAP.get(c, {}).get(v, {}).keys()), cam, list(DATA_MAP.get(c, {}).get(v, {}).get(cam, {}).keys()), s, DATA_MAP.get(c, {}).get(v, {}).get(cam, {}).get(s, []), o
+
+CHUNKS, C, VIDS, V, CAMS, CAM, SHOTS, S, OBJS, O = get_init()
+
+def get_choices_and_val(lst):
+    return gr.update(choices=lst, value=lst[0] if lst else "-")
+
+def get_valid_replace_shots(chunk, vid, cam, obj):
+    if "-" in [chunk, vid, cam, obj]: return []
+    return [s for s in list(DATA_MAP.get(chunk, {}).get(vid, {}).get(cam, {}).keys()) if obj in DATA_MAP.get(chunk, {}).get(vid, {}).get(cam, {}).get(s, [])]
+
+def chunk_update(chunk):
+    vids = list(DATA_MAP.get(chunk, {}).keys())
+    cams = list(DATA_MAP.get(chunk, {}).get(vids[0] if vids else "-", {}).keys())
+    shots = list(DATA_MAP.get(chunk, {}).get(vids[0] if vids else "-", {}).get(cams[0] if cams else "-", {}).keys())
+    objs = DATA_MAP.get(chunk, {}).get(vids[0] if vids else "-", {}).get(cams[0] if cams else "-", {}).get(shots[0] if shots else "-", [])
+    vr = get_valid_replace_shots(chunk, vids[0] if vids else "-", cams[0] if cams else "-", objs[0] if objs else "-")
+    return get_choices_and_val(vids), get_choices_and_val(cams), get_choices_and_val(shots), get_choices_and_val(objs), gr.update(choices=vr, value=vr, label=f"使用 [{objs[0] if objs else '-'}] 擦除版的片段")
+
+def vid_update(chunk, vid):
+    cams = list(DATA_MAP.get(chunk, {}).get(vid, {}).keys())
+    shots = list(DATA_MAP.get(chunk, {}).get(vid, {}).get(cams[0] if cams else "-", {}).keys())
+    objs = DATA_MAP.get(chunk, {}).get(vid, {}).get(cams[0] if cams else "-", {}).get(shots[0] if shots else "-", [])
+    vr = get_valid_replace_shots(chunk, vid, cams[0] if cams else "-", objs[0] if objs else "-")
+    return get_choices_and_val(cams), get_choices_and_val(shots), get_choices_and_val(objs), gr.update(choices=vr, value=vr, label=f"使用 [{objs[0] if objs else '-'}] 擦除版的片段")
+
+def cam_update(chunk, vid, cam):
+    shots = list(DATA_MAP.get(chunk, {}).get(vid, {}).get(cam, {}).keys())
+    objs = DATA_MAP.get(chunk, {}).get(vid, {}).get(cam, {}).get(shots[0] if shots else "-", [])
+    vr = get_valid_replace_shots(chunk, vid, cam, objs[0] if objs else "-")
+    return get_choices_and_val(shots), get_choices_and_val(objs), gr.update(choices=vr, value=vr, label=f"使用 [{objs[0] if objs else '-'}] 擦除版的片段")
+
+def shot_update(chunk, vid, cam, shot):
+    objs = DATA_MAP.get(chunk, {}).get(vid, {}).get(cam, {}).get(shot, [])
+    vr = get_valid_replace_shots(chunk, vid, cam, objs[0] if objs else "-")
+    return get_choices_and_val(objs), gr.update(choices=vr, value=vr, label=f"使用 [{objs[0] if objs else '-'}] 擦除版的片段")
+
+def obj_update(chunk, vid, cam, obj):
+    vr = get_valid_replace_shots(chunk, vid, cam, obj)
+    return gr.update(choices=vr, value=vr, label=f"使用 [{obj}] 擦除版的片段")
+
+def create_overlay_video(orig_path, mask_dir, out_path):
+    if os.path.exists(out_path): return out_path
+    
+    cap = cv2.VideoCapture(orig_path)
+    fps = cap.get(cv2.CAP_PROP_FPS) or 24
+    orig_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    orig_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    
+    # 【新增优化】设置缩放比例为 0.5（长宽各减半）
+    scale = 0.5
+    # 【新增优化】确保新计算出的长宽必须是偶数 (H.264 编码器的硬性要求，否则会报错)
+    w = int((orig_w * scale) // 2 * 2)
+    h = int((orig_h * scale) // 2 * 2)
+
+    mask_files = sorted(glob.glob(os.path.join(mask_dir, "*.png")))
+    temp_path = out_path.replace('.mp4', '_temp.mp4')
+    out = cv2.VideoWriter(temp_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+    
+    idx = 0
+    while True:
+        ret, frame = cap.read()
+        if not ret: break
+        
+        # 【新增优化】把原始视频帧缩小
+        frame = cv2.resize(frame, (w, h), interpolation=cv2.INTER_AREA)
+        
+        if idx < len(mask_files):
+            mask = cv2.imread(mask_files[idx], cv2.IMREAD_GRAYSCALE)
+            if mask is not None:
+                # 把 Mask 也缩放到相同的小尺寸
+                mask = cv2.resize(mask, (w, h), interpolation=cv2.INTER_NEAREST)
+                frame[mask > 128] = frame[mask > 128] * 0.5 + np.array([0, 0, 255], dtype=np.uint8) * 0.5
+        
+        out.write(frame)
+        idx += 1
+        
+    cap.release()
+    out.release()
+    
+    # 【优化参数】将 -crf 从 28 提高到 32（牺牲一点画质换取更快的编码速度）
+    # 注意这里默认你已经用了 FFMPEG_PATH 变量
+    subprocess.run([FFMPEG_PATH, '-y', '-i', temp_path, '-c:v', 'libx264', '-preset', 'ultrafast', '-crf', '32', out_path], capture_output=True)
+    
+    if os.path.exists(temp_path): os.remove(temp_path)
+    return out_path
+def load_preview(chunk, vid, cam, shot, obj):
+    if not all([chunk, vid, cam, shot, obj]) or "-" in [chunk, vid, cam, shot, obj]: return None, None
+    orig = os.path.join(CI_VID_DIR, chunk, vid, cam, f"{shot}.mp4")
+    edit = os.path.join(STAGE4_DIR, chunk, vid, cam, str(shot), f"{obj}.mp4")
+    mask_dir = os.path.join(STAGE3_DIR, chunk, vid, cam, str(shot), str(obj))
+    orig_to_show = create_overlay_video(orig, mask_dir, os.path.join(CACHE_DIR, f"ov_{chunk}_{vid}_{cam}_{shot}_{obj}.mp4")) if os.path.exists(orig) and os.path.exists(mask_dir) else (orig if os.path.exists(orig) else None)
+    return orig_to_show, (edit if os.path.exists(edit) else None)
+
+def merge_videos(chunk, vid, cam, obj, replace_shots):
+    if not all([chunk, vid, cam, obj]) or "-" in [chunk, vid, cam, obj]: return None
+    all_orig = sorted([p for p in glob.glob(os.path.join(CI_VID_DIR, chunk, vid, cam, "*.mp4")) if os.path.basename(p).replace('.mp4','').isdigit()], key=lambda x: int(os.path.basename(x).split('.')[0]))
+    out_path = os.path.join(CACHE_DIR, f"merged_{chunk}_{vid}_{cam}_{obj}.mp4")
+    inputs = []
+    for p in all_orig:
+        sid = os.path.basename(p).replace('.mp4','')
+        edit_p = os.path.join(STAGE4_DIR, chunk, vid, cam, sid, f"{obj}.mp4")
+        inputs.append(edit_p if (sid in replace_shots and os.path.exists(edit_p)) else p)
+    if not inputs: return None
+    flt = "".join([f"[{i}:v]scale=1280:720:force_original_aspect_ratio=decrease,pad=1280:720:(ow-iw)/2:(oh-ih)/2,setsar=1,fps=24[v{i}];" for i in range(len(inputs))])
+    concat = "".join([f"[v{i}]" for i in range(len(inputs))]) + f"concat=n={len(inputs)}:v=1:a=0[outv]"
+    cmd = [FFMPEG_PATH, "-y", *sum([["-i", inp] for inp in inputs], []), "-filter_complex", flt + concat, "-map", "[outv]", "-c:v", "libx264", "-preset", "ultrafast", out_path]
+    subprocess.run(cmd, check=True, capture_output=True)
+    return out_path
+
+with gr.Blocks() as app:
+    gr.Markdown("# 🎬 视频不一致性检查面板 (Multi-Object版)")
+    with gr.Row():
+        chunk_dd = gr.Dropdown(choices=CHUNKS, label="1. Chunk", value=C)
+        vid_dd   = gr.Dropdown(choices=VIDS, label="2. Video ID", value=V)
+        cam_dd   = gr.Dropdown(choices=CAMS, label="3. Camera ID", value=CAM)
+    gr.Markdown("---")
+    with gr.Row():
+        with gr.Column():
+            with gr.Row():
+                shot_dd = gr.Dropdown(choices=SHOTS, label="4. Shot", value=S)
+                obj_dd  = gr.Dropdown(choices=OBJS, label="5. Object", value=O)
+            v_orig = gr.Video(label="原始片段 (已叠加红色Mask)")
+            v_edit = gr.Video(label="擦除后结果")
+        with gr.Column():
+            replace = gr.CheckboxGroup(choices=get_valid_replace_shots(C, V, CAM, O), label=f"使用 [{O}] 擦除版的片段", value=get_valid_replace_shots(C, V, CAM, O))
+            btn     = gr.Button("🚀 一键合成预览视频", variant="primary")
+            v_res   = gr.Video(label="全序列拼接结果")
+
+    chunk_dd.change(chunk_update, [chunk_dd], [vid_dd, cam_dd, shot_dd, obj_dd, replace])
+    vid_dd.change(vid_update, [chunk_dd, vid_dd], [cam_dd, shot_dd, obj_dd, replace])
+    cam_dd.change(cam_update, [chunk_dd, vid_dd, cam_dd], [shot_dd, obj_dd, replace])
+    shot_dd.change(shot_update, [chunk_dd, vid_dd, cam_dd, shot_dd], [obj_dd, replace])
+    obj_dd.change(obj_update, [chunk_dd, vid_dd, cam_dd, obj_dd], [replace])
+    
+    shot_dd.change(load_preview, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [v_orig, v_edit])
+    obj_dd.change(load_preview, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [v_orig, v_edit])
+    btn.click(merge_videos, [chunk_dd, vid_dd, cam_dd, obj_dd, replace], [v_res])
+    app.load(load_preview, [chunk_dd, vid_dd, cam_dd, shot_dd, obj_dd], [v_orig, v_edit])
+
+if __name__ == "__main__":
+    print(f"Link: http://127.0.0.1:7860/")
+    app.launch(server_name="0.0.0.0", server_port=7860, share=True, allowed_paths=["/workspace"])