Object detect/yeyoung

config.py

@@ -0,0 +1,47 @@
+# config.py
+
+from pathlib import Path
+import torch
+
+# =========================
+# Project Paths
+# =========================
+PROJECT_ROOT = Path(__file__).resolve().parent
+
+OUTPUT_DIR = PROJECT_ROOT / "outputs"
+OBJECT_OUTPUT_DIR = OUTPUT_DIR / "object"
+
+OBJECT_DB_PATH = OBJECT_OUTPUT_DIR / "object_db.json"
+OBJECT_DEBUG_CROP_DIR = OBJECT_OUTPUT_DIR / "debug_crops"
+
+# =========================
+# Device
+# =========================
+if torch.backends.mps.is_available():
+    DEVICE = "mps"
+elif torch.cuda.is_available():
+    DEVICE = "cuda"
+else:
+    DEVICE = "cpu"
+
+# =========================
+# Object Detection / Privacy Reasoning
+# =========================
+OBJECT_QWEN_MODEL_NAME = "Qwen/Qwen2-VL-2B-Instruct"
+OBJECT_DEFAULT_PROMPT = "내 프라이버시가 유출될 만한 것들을 가려줘."
+OBJECT_SAMPLE_FPS = 1.0
+OBJECT_TEXT_DETECTOR_LANG = "korean"
+
+OBJECT_CROP_MARGIN_RATIO = 0.35
+OBJECT_MAX_NEW_TOKENS_REASON = 48
+OBJECT_TRACK_IOU_THRESHOLD = 0.30
+
+OBJECT_MIN_TEXT_BOX_AREA = 16
+OBJECT_MAX_TEXT_BOX_AREA_RATIO = 0.75
+OBJECT_MIN_TEXT_REGION_WIDTH = 3
+OBJECT_MIN_TEXT_REGION_HEIGHT = 3
+OBJECT_MIN_REC_SCORE = 0.30
+OBJECT_MIN_GROUP_ITEMS = 1
+OBJECT_GROUP_X_GAP_RATIO = 1.6
+OBJECT_GROUP_Y_GAP_RATIO = 1.0
+OBJECT_MIN_QWEN_CROP_SIZE = 56

main.py

@@ -0,0 +1,58 @@
+# main.py
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+import config
+from pipeline.pass3_object_detect import run_pass3
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Privacy Blur Pipeline - Object Detection Test Entry"
+    )
+
+    parser.add_argument(
+        "--video",
+        required=True,
+        help="입력 영상 경로",
+    )
+
+    parser.add_argument(
+        "--prompt",
+        default=config.OBJECT_DEFAULT_PROMPT,
+        help="객체 탐지용 자연어 프롬프트",
+    )
+
+    parser.add_argument(
+        "--sample-fps",
+        type=float,
+        default=config.OBJECT_SAMPLE_FPS,
+        help="객체 탐지 샘플링 FPS",
+    )
+
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+
+    video_path = Path(args.video).expanduser().resolve()
+
+    if not video_path.exists():
+        raise FileNotFoundError(f"입력 영상을 찾을 수 없습니다: {video_path}")
+
+    config.OBJECT_OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
+
+    run_pass3(
+        video_path=video_path,
+        user_prompt=args.prompt,
+        sample_fps=args.sample_fps,
+        output_dir=config.OBJECT_OUTPUT_DIR,
+    )
+
+
+if __name__ == "__main__":
+    main()