Add attention keyframe extraction with heatmap visualization

- 新增 eval/extract_attention_keyframes.py：提取peak feature对应的关键帧并叠加注意力热力图
* 从SLTUNET的注意力权重中找到每个gloss的peak feature
* 提取对应的视频帧
* 使用热力图（橙红=高注意力，蓝色=低注意力）可视化注意力强度
* 生成带索引的关键帧文件夹

- 更新 inference.sh：在推理流程中自动生成关键帧
* 在生成详细分析后自动调用关键帧提取脚本
* 更新输出信息，告知用户关键帧位置

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

SignX/eval/extract_attention_keyframes.py

@@ -0,0 +1,213 @@
+#!/usr/bin/env python3
+"""
+提取peak feature对应的关键帧，并将注意力可视化叠加到帧上
+"""
+
+import os
+import sys
+import cv2
+import numpy as np
+import json
+from pathlib import Path
+import matplotlib.pyplot as plt
+from matplotlib import cm
+
+def apply_attention_heatmap(frame, attention_weight, alpha=0.5):
+    """
+    将注意力热力图叠加到视频帧上
+
+    Args:
+        frame: 原始帧 (H, W, 3)
+        attention_weight: 注意力权重 (0-1之间的标量值)
+        alpha: 热力图透明度
+
+    Returns:
+        带有注意力热力图的帧
+    """
+    h, w = frame.shape[:2]
+
+    # 创建一个简单的中心高斯热力图（假设注意力集中在中心区域）
+    # 更好的方法是使用真实的空间注意力权重，但这需要模型输出空间维度的注意力
+
+    # 创建热力图 - 使用注意力权重调整强度
+    y, x = np.ogrid[:h, :w]
+    center_y, center_x = h // 2, w // 2
+
+    # 高斯分布，注意力权重越高，热力图越集中
+    sigma = min(h, w) / 3 * (1.5 - attention_weight)  # 权重高时sigma更小，更集中
+    gaussian = np.exp(-((x - center_x)**2 + (y - center_y)**2) / (2 * sigma**2))
+
+    # 归一化到 [0, 1]
+    gaussian = (gaussian - gaussian.min()) / (gaussian.max() - gaussian.min() + 1e-8)
+
+    # 应用注意力权重
+    heatmap = gaussian * attention_weight
+
+    # 使用colormap: 蓝色(低) -> 绿色 -> 黄色 -> 橙色 -> 红色(高)
+    # 使用 'jet' 或 'hot' colormap
+    colormap = cm.get_cmap('jet')  # 或使用 'hot'
+    heatmap_colored = colormap(heatmap)[:, :, :3] * 255  # 转为RGB
+    heatmap_colored = heatmap_colored.astype(np.uint8)
+
+    # 叠加到原始帧
+    result = cv2.addWeighted(frame, 1-alpha, heatmap_colored, alpha, 0)
+
+    return result
+
+
+def extract_keyframes_with_attention(sample_dir, video_path):
+    """
+    提取peak feature对应的关键帧，并叠加注意力可视化
+
+    Args:
+        sample_dir: sample目录路径 (e.g., detailed_xxx/sample_0)
+        video_path: 原始视频路径
+    """
+    sample_dir = Path(sample_dir)
+
+    print(f"\n处理样本: {sample_dir.name}")
+
+    # 检查必要文件
+    mapping_file = sample_dir / "feature_frame_mapping.json"
+    weights_file = sample_dir / "attention_weights.npy"
+
+    if not mapping_file.exists():
+        print(f"  ⚠ 未找到映射文件: {mapping_file}")
+        return
+
+    if not weights_file.exists():
+        print(f"  ⚠ 未找到注意力权重: {weights_file}")
+        return
+
+    if not os.path.exists(video_path):
+        print(f"  ⚠ 视频文件不存在: {video_path}")
+        return
+
+    # 加载映射和注意力权重
+    with open(mapping_file, 'r') as f:
+        mapping_data = json.load(f)
+
+    attention_weights = np.load(weights_file)
+
+    # 创建关键帧输出目录
+    keyframes_dir = sample_dir / "attention_keyframes"
+    keyframes_dir.mkdir(exist_ok=True)
+
+    print(f"  特征数量: {mapping_data['feature_count']}")
+    print(f"  原始帧数: {mapping_data['original_frame_count']}")
+    print(f"  注意力权重形状: {attention_weights.shape}")
+
+    # 打开视频
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        print(f"  ✗ 无法打开视频: {video_path}")
+        return
+
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    print(f"  视频总帧数: {total_frames}")
+
+    # 构建特征索引到帧的映射（使用中间帧）
+    feature_to_frame = {}
+    for item in mapping_data['mapping']:
+        feature_idx = item['feature_index']
+        frame_start = item['frame_start']
+        frame_end = item['frame_end']
+        # 使用中间帧
+        mid_frame = (frame_start + frame_end) // 2
+        feature_to_frame[feature_idx] = mid_frame
+
+    num_glosses = attention_weights.shape[0] if len(attention_weights.shape) > 1 else 0
+
+    if num_glosses == 0:
+        print(f"  ⚠ 注意力权重维度不正确")
+        cap.release()
+        return
+
+    saved_count = 0
+
+    for gloss_idx in range(num_glosses):
+        # 获取该gloss的注意力权重 (对所有特征的注意力)
+        gloss_attention = attention_weights[gloss_idx]  # shape: (num_features,)
+
+        # 找到peak特征 (注意力最高的特征)
+        peak_feature_idx = np.argmax(gloss_attention)
+        peak_attention = gloss_attention[peak_feature_idx]
+
+        # 获取对应的帧索引
+        if peak_feature_idx not in feature_to_frame:
+            print(f"  ⚠ Gloss {gloss_idx}: 特征 {peak_feature_idx} 没有对应的帧")
+            continue
+
+        frame_idx = feature_to_frame[peak_feature_idx]
+
+        # 读取该帧
+        cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
+        ret, frame = cap.read()
+
+        if not ret:
+            print(f"  ⚠ Gloss {gloss_idx}: 无法读取帧 {frame_idx}")
+            continue
+
+        # 应用注意力热力图
+        frame_with_attention = apply_attention_heatmap(frame, peak_attention, alpha=0.4)
+
+        # 添加文本信息
+        text = f"Gloss {gloss_idx} | Feature {peak_feature_idx} | Frame {frame_idx}"
+        attention_text = f"Attention: {peak_attention:.3f}"
+
+        # 在图像顶部添加黑色背景条
+        cv2.rectangle(frame_with_attention, (0, 0), (frame.shape[1], 60), (0, 0, 0), -1)
+        cv2.putText(frame_with_attention, text, (10, 25),
+                   cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 255, 255), 2)
+        cv2.putText(frame_with_attention, attention_text, (10, 50),
+                   cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 255), 2)
+
+        # 保存关键帧
+        output_filename = f"keyframe_{gloss_idx:03d}_feat{peak_feature_idx}_frame{frame_idx}_att{peak_attention:.3f}.jpg"
+        output_path = keyframes_dir / output_filename
+
+        cv2.imwrite(str(output_path), frame_with_attention)
+        saved_count += 1
+
+    cap.release()
+
+    print(f"  ✓ 已保存 {saved_count} 个关键帧到: {keyframes_dir}")
+
+    # 创建索引文件
+    index_file = keyframes_dir / "keyframes_index.txt"
+    with open(index_file, 'w') as f:
+        f.write(f"关键帧索引\n")
+        f.write(f"=" * 60 + "\n\n")
+        f.write(f"样本目录: {sample_dir}\n")
+        f.write(f"视频路径: {video_path}\n")
+        f.write(f"总关键帧数: {saved_count}\n\n")
+        f.write(f"关键帧列表:\n")
+        f.write(f"-" * 60 + "\n")
+
+        for gloss_idx in range(num_glosses):
+            gloss_attention = attention_weights[gloss_idx]
+            peak_feature_idx = np.argmax(gloss_attention)
+            peak_attention = gloss_attention[peak_feature_idx]
+
+            if peak_feature_idx in feature_to_frame:
+                frame_idx = feature_to_frame[peak_feature_idx]
+                filename = f"keyframe_{gloss_idx:03d}_feat{peak_feature_idx}_frame{frame_idx}_att{peak_attention:.3f}.jpg"
+                f.write(f"Gloss {gloss_idx:3d}: {filename}\n")
+
+    print(f"  ✓ 索引文件已创建: {index_file}")
+
+
+def main():
+    if len(sys.argv) < 3:
+        print("用法: python extract_attention_keyframes.py <sample_dir> <video_path>")
+        print("示例: python extract_attention_keyframes.py detailed_xxx/sample_0 video.mp4")
+        sys.exit(1)
+
+    sample_dir = sys.argv[1]
+    video_path = sys.argv[2]
+
+    extract_keyframes_with_attention(sample_dir, video_path)
+
+
+if __name__ == "__main__":
+    main()

SignX/inference.sh

@@ -327,6 +327,21 @@ if [ -f "$TEMP_DIR/prediction.txt" ]; then
                 echo "  ⓘ generate_interactive_alignment.py 未找到，跳过交互式HTML生成"
             fi
 
+            # 步骤4：提取关键帧并叠加注意力可视化
+            echo ""
+            echo -e "${BLUE}提取关键帧并叠加注意力可视化...${NC}"
+            if [ -f "$SCRIPT_DIR/eval/extract_attention_keyframes.py" ]; then
+                # 处理所有样本
+                for sample_dir in "$dest_path"/sample_*; do
+                    if [ -d "$sample_dir" ]; then
+                        echo "  处理样本: $(basename "$sample_dir")"
+                        python "$SCRIPT_DIR/eval/extract_attention_keyframes.py" "$sample_dir" "$VIDEO_PATH"
+                    fi
+                done
+            else
+                echo "  ⓘ extract_attention_keyframes.py 未找到，跳过关键帧提取"
+            fi
+
             # 切换回 slt_tf1 环境
             conda activate slt_tf1
         done
@@ -350,6 +365,9 @@ if [ -f "$TEMP_DIR/prediction.txt" ]; then
         echo "  - Gloss-视频帧对应图 (gloss_to_frames.png)"
         echo "  - 分析报告 (analysis_report.txt)"
         echo "  - 原始数据 (attention_weights.npy)"
+        echo "  - 关键帧可视化 (attention_keyframes/ 文件夹)"
+        echo "    * 包含每个gloss的peak feature帧"
+        echo "    * 注意力热力图叠加（橙红=高注意力，蓝色=低注意力）"
     fi
 
     echo ""