{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [markdown]\n", "# # [1] 초기 설정 & 라이브러리 임포트\n", "\n", "# %%\n", "import os\n", "import sys\n", "import cv2\n", "import numpy as np\n", "import pandas as pd\n", "import matplotlib.pyplot as plt\n", "\n", "import torch\n", "\n", "from mmengine.registry import init_default_scope\n", "from mmdet.apis import init_detector, inference_detector\n", "from mmdet.utils import register_all_modules\n", "\n", "# 필요 시 mmyolo path 추가\n", "mmyolo_dir = \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo\"\n", "if mmyolo_dir not in sys.path:\n", " sys.path.append(mmyolo_dir)\n", "\n", "# 레지스트리 등록\n", "register_all_modules()\n", "init_default_scope('mmdet')\n", "\n", "# ----------------------------------------------------\n", "# (!!!) 사용자 지정 파라미터\n", "# ----------------------------------------------------\n", "VIDEO_PATH = \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/val_sample/포스코_광주_고양_합침/KakaoTalk_20250108_154050505.mp4\"\n", "MAIN_OUTPUT_DIR = \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/val_sample/모델비교_아웃풋들\"\n", "device = \"cuda:0\" if torch.cuda.is_available() else \"cpu\"\n", "print(f\"🚀 디바이스: {device}\")\n", "\n", "# 4개 모델 정보\n", "MODEL_INFOS = [\n", " {\n", " \"model_name\": \"1819crc\",\n", " \"config\": \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo/work_dirs/ff_base_1819+cut+rotate+cutrota_1,1,3_01/yolo8_large_960size_ff_base_1819+cut+rotate+cutrota_1,1,3.py\",\n", " \"checkpoint\": \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo/work_dirs/ff_base_1819+cut+rotate+cutrota_1,1,3_01/best_coco_bbox_mAP_epoch_6.pth\",\n", " \"class_names\": [\"worker\", \"helmet\", \"head\"],\n", " \"score_thresholds\": {\"worker\":0.1, \"helmet\":0.1, \"head\":0.1}\n", " },\n", " {\n", " \"model_name\": \"resize_1819crc\",\n", " \"config\": \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo/work_dirs/ff_base_resize_1,2,0.5_1819+cut+rotate+cutrota_1,1,3_01/yolo8_large_960size_ff_base_resize_1,2,0.5_1819+cut+rotate+cutrota_1,1,3.py\",\n", " \"checkpoint\": \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo/work_dirs/ff_base_resize_1,2,0.5_1819+cut+rotate+cutrota_1,1,3_01/best_coco_bbox_mAP_epoch_8.pth\",\n", " \"class_names\": [\"worker\", \"helmet\", \"head\"],\n", " \"score_thresholds\": {\"worker\":0.1, \"helmet\":0.1, \"head\":0.1}\n", " },\n", " {\n", " \"model_name\": \"load_1819\",\n", " \"config\": \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo/work_dirs/it1_load_epoch16_1819_add_head_2,1,10_01/yolov8_l_960_custom_add_head_ff_result_2,1,10.py\",\n", " \"checkpoint\": \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo/work_dirs/it1_load_epoch16_1819_add_head_2,1,10_01/best_coco_bbox_mAP_epoch_19.pth\",\n", " \"class_names\": [\"worker\", \"helmet\", \"head\"],\n", " \"score_thresholds\": {\"worker\":0.1, \"helmet\":0.1, \"head\":0.1}\n", " },\n", " {\n", " \"model_name\": \"fall_960\",\n", " \"config\": \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo/work_dirs/it1_roboflow_fall_only_rotate_960_yolo_01/yolov8_l_960_custom_fall.py\",\n", " \"checkpoint\": \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/mmyolo/work_dirs/it1_roboflow_fall_only_rotate_960_yolo_01/best_coco_bbox_mAP_epoch_15.pth\",\n", " \"class_names\": [\"worker\"], # fall 모델은 1클래스만\n", " \"score_thresholds\": {\"worker\":0.1}\n", " },\n", "]\n", "\n", "# 클래스별 색상 매핑 (필요시 수정)\n", "CLASS_COLORS = {\n", " \"worker\": (0, 255, 0), # 초록\n", " \"helmet\": (255, 0, 0), # 파랑\n", " \"head\": (0, 0, 255), # 빨강\n", " # 그 외 클래스는 기본 회색\n", "}\n", "\n", "print(\"✅ [1] 초기 설정 완료.\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [markdown]\n", "# # [2] 결과 폴더 생성 함수\n", "# \n", "# - exp, exp2... 식으로 저장 폴더 생성해, 중복 방지.\n", "\n", "# %%\n", "def get_new_output_folder(base_dir, prefix=\"exp\"):\n", " i = 1\n", " while True:\n", " folder_name = prefix if i == 1 else f\"{prefix}{i}\"\n", " new_dir = os.path.join(base_dir, folder_name)\n", " if not os.path.exists(new_dir):\n", " os.makedirs(new_dir)\n", " return new_dir\n", " i += 1\n", "\n", "output_folder = get_new_output_folder(MAIN_OUTPUT_DIR, prefix=\"exp\")\n", "print(f\"✅ 결과 저장 폴더: {output_folder}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [markdown]\n", "# # [3] 모델 로드 & 디텍션 함수\n", "# \n", "# - init_detector()\n", "# - detect_objects() -> \"worker\"만 남긴다\n", "# - draw_detections_advanced() -> bbox 옆에 score, w,h,area 표시\n", "\n", "# %%\n", "def load_model(cfg, ckp, device=\"cpu\"):\n", " model = init_detector(cfg, ckp, device=device)\n", " return model\n", "\n", "def detect_objects(frame_bgr, model, class_names, class_score_thresholds):\n", " \"\"\"\n", " -> \"worker\" 클래스만 유지. helmet/head는 무시.\n", " \"\"\"\n", " result = inference_detector(model, frame_bgr)\n", " if not hasattr(result, \"pred_instances\"):\n", " return []\n", " \n", " bboxes = result.pred_instances.bboxes\n", " scores = result.pred_instances.scores\n", " labels = result.pred_instances.labels\n", " \n", " out = []\n", " for i in range(len(bboxes)):\n", " box = bboxes[i].tolist() # [x1,y1,x2,y2]\n", " sc = scores[i].item()\n", " cid = labels[i].item()\n", " # class 이름\n", " if cid < len(class_names):\n", " cname = class_names[cid]\n", " else:\n", " cname = f\"cls_{cid}\"\n", " \n", " # 여기서 \"worker\"만 필터\n", " if cname != \"worker\":\n", " continue\n", " \n", " thr = class_score_thresholds.get(\"worker\", 0.1)\n", " if sc < thr:\n", " continue\n", " \n", " x1, y1, x2, y2 = box\n", " w = x2 - x1\n", " h = y2 - y1\n", " area = w*h\n", " out.append({\n", " \"bbox\":[x1,y1,x2,y2],\n", " \"score\": sc,\n", " \"class_name\": \"worker\", # 고정\n", " \"width\": w,\n", " \"height\":h,\n", " \"area\": area\n", " })\n", " return out\n", "\n", "def draw_detections_advanced(frame_bgr, detections, class_colors, font_scale=1.0, font_thick=2):\n", " \"\"\"\n", " 조금 더 크게 표시. (font_scale=1.0)\n", " \"\"\"\n", " result = frame_bgr.copy()\n", " for det in detections:\n", " x1,y1,x2,y2 = map(int, det[\"bbox\"])\n", " sc = det[\"score\"]\n", " w = det[\"width\"]\n", " h = det[\"height\"]\n", " area = det[\"area\"]\n", " cname = det[\"class_name\"]\n", " \n", " color = class_colors.get(cname, (128,128,128))\n", " \n", " # bbox\n", " cv2.rectangle(result, (x1,y1), (x2,y2), color, 3)\n", " # 텍스트\n", " line1 = f\"{cname} {sc:.2f}\"\n", " line2 = f\"W={w:.0f},H={h:.0f},A={area:.0f}\"\n", " \n", " # 위치 계산\n", " y_text = y1-10\n", " cv2.putText(result, line1, (x1, y_text),\n", " cv2.FONT_HERSHEY_SIMPLEX, font_scale, color, font_thick, cv2.LINE_AA)\n", " y_text -= 35\n", " cv2.putText(result, line2, (x1, max(y_text,10)),\n", " cv2.FONT_HERSHEY_SIMPLEX, font_scale, color, font_thick, cv2.LINE_AA)\n", " \n", " return result\n", "\n", "print(\"✅ [3] 모델 로드 & 디텍션 함수 준비 완료.\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [markdown]\n", "# # [4] 부분 프레임 추론 & 2x2 합치기\n", "\n", "# %%\n", "def make_image_grid(img_list, rows=2, cols=2):\n", " \"\"\"\n", " 2x2로 합친 뒤 반환. \n", " img_list: 최대 4장. (4개 이하라면 blank로 채움)\n", " \"\"\"\n", " if not img_list:\n", " return None\n", " # 만약 4개보다 적으면 blank 채우기\n", " base = img_list[0]\n", " h, w = base.shape[:2]\n", " blank = np.zeros_like(base)\n", " \n", " while len(img_list) < rows*cols:\n", " img_list.append(blank)\n", " \n", " grid = np.zeros((h*rows, w*cols, 3), dtype=np.uint8)\n", " idx=0\n", " for r in range(rows):\n", " for c in range(cols):\n", " grid[r*h:(r+1)*h, c*w:(c+1)*w] = img_list[idx]\n", " idx+=1\n", " return grid\n", "\n", "def run_detection_for_time(video_path, time_in_seconds, loaded_models,\n", " frame_range=2, \n", " save_dir=None):\n", " \"\"\"\n", " - time_in_seconds -> frame_index\n", " - frame_index ± frame_range 추론\n", " - 'worker'만 디텍션 \n", " - 4개 모델 결과를 2x2로 합친 뒤, 해상도 2배로 resize하여 저장\n", " - 리턴: det_records\n", " \"\"\"\n", " det_records = []\n", " \n", " cap = cv2.VideoCapture(video_path)\n", " if not cap.isOpened():\n", " print(f\"[에러] 영상 열기 실패: {video_path}\")\n", " return det_records\n", " \n", " fps = cap.get(cv2.CAP_PROP_FPS)\n", " total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))\n", " \n", " targ_f = int(time_in_seconds * fps)\n", " start_f = max(0, targ_f - frame_range)\n", " end_f = min(total_frames-1, targ_f + frame_range)\n", " \n", " print(f\"🎯 초={time_in_seconds}, frame_range=±{frame_range}\")\n", " print(f\" -> 프레임 구간: [{start_f} ~ {end_f}]\")\n", " \n", " for fidx in range(start_f, end_f+1):\n", " cap.set(cv2.CAP_PROP_POS_FRAMES, fidx)\n", " ret, frame_bgr = cap.read()\n", " if not ret:\n", " print(f\"[경고] 프레임 {fidx} 읽기 실패.\")\n", " continue\n", " \n", " model_vis_list = []\n", " model_summaries = []\n", " \n", " for m_info in loaded_models:\n", " m_name = m_info[\"model_name\"]\n", " m_obj = m_info[\"model_obj\"]\n", " cls_n = m_info[\"class_names\"]\n", " thr_d = m_info[\"score_thresholds\"]\n", " \n", " # detect\n", " dets = detect_objects(frame_bgr, m_obj, cls_n, thr_d) # worker만\n", " # draw\n", " vis = draw_detections_advanced(frame_bgr, dets, CLASS_COLORS,\n", " font_scale=1.0, font_thick=2)\n", " \n", " model_vis_list.append(vis)\n", " \n", " # record\n", " for d in dets:\n", " rec = {\n", " \"model_name\": m_name,\n", " \"frame_idx\": fidx,\n", " \"score\": d[\"score\"],\n", " \"width\": d[\"width\"],\n", " \"height\": d[\"height\"],\n", " \"area\": d[\"area\"]\n", " }\n", " det_records.append(rec)\n", " \n", " # 요약\n", " n_det = len(dets)\n", " mean_sc = np.mean([d[\"score\"] for d in dets]) if n_det>0 else 0\n", " mean_ar = np.mean([d[\"area\"] for d in dets]) if n_det>0 else 0\n", " model_summaries.append((m_name, n_det, mean_sc, mean_ar))\n", " \n", " if model_vis_list:\n", " grid_img = make_image_grid(model_vis_list,2,2)\n", " \n", " # 해상도 2배로 키우기\n", " gh, gw = grid_img.shape[:2]\n", " # 배율. 여기서는 1.5배 or 2배 (자유롭게)\n", " scale_factor = 2.0\n", " grid_resized = cv2.resize(grid_img, (int(gw*scale_factor), int(gh*scale_factor)))\n", " \n", " # 상단 요약 텍스트\n", " summ_text = f\"Frame {fidx}\\n\"\n", " for (m_name, n_det, m_sc, m_ar) in model_summaries:\n", " summ_text += f\"{m_name}: #Wkr={n_det}, avgSc={m_sc:.2f}, avgArea={m_ar:.1f}\\n\"\n", " \n", " # 텍스트 표시 (왼쪽 위에)\n", " y_off = 40\n", " for line in summ_text.split(\"\\n\"):\n", " cv2.putText(grid_resized, line, (20,y_off),\n", " cv2.FONT_HERSHEY_SIMPLEX, 1.2, (255,255,255), 3, cv2.LINE_AA)\n", " y_off += 40\n", " \n", " if save_dir:\n", " base_nm = os.path.splitext(os.path.basename(video_path))[0]\n", " out_fn = f\"{base_nm}_frame{fidx:04d}_4model_grid.jpg\"\n", " out_p = os.path.join(save_dir, out_fn)\n", " \n", " cv2.imwrite(out_p, grid_resized)\n", " plt.figure(figsize=(12,10))\n", " plt.imshow(cv2.cvtColor(grid_resized, cv2.COLOR_BGR2RGB))\n", " plt.title(f\"Frame {fidx} - Worker Only\")\n", " plt.axis(\"off\")\n", " plt.show()\n", " \n", " cap.release()\n", " return det_records\n", "\n", "print(\"✅ [4] 부분 프레임 추론 & 합치기 함수 완료.\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [markdown]\n", "# # [5] 실행 예시\n", "# \n", "# - 모델들 로드\n", "# - 특정 초(예: 10초) ± 2프레임\n", "# - 결과 이미지 + df_det\n", "\n", "# %%\n", "loaded_models = []\n", "for info in MODEL_INFOS:\n", " print(f\"🔄 로딩: {info['model_name']}\")\n", " mdl = load_model(info[\"config\"], info[\"checkpoint\"], device=device)\n", " loaded_models.append({\n", " \"model_name\": info[\"model_name\"],\n", " \"model_obj\": mdl,\n", " \"class_names\": info[\"class_names\"],\n", " \"score_thresholds\": info[\"score_thresholds\"]\n", " })\n", "\n", "time_in_sec = 26.0\n", "frame_range = 1\n", "\n", "det_list = run_detection_for_time(VIDEO_PATH, time_in_sec, loaded_models,\n", " frame_range=frame_range,\n", " save_dir=output_folder)\n", "df_det = pd.DataFrame(det_list)\n", "print(\"df_det.shape:\", df_det.shape)\n", "print(df_det.head(10))\n", "\n", "print(\"✅ [5] 실행 예시 완료.\")\n", "print(f\"결과 폴더: {output_folder}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [markdown]\n", "# # [6] 프레임별 모델별 분석 & 그래프\n", "# \n", "# - df_det: columns=[model_name, frame_idx, score, width, height, area]\n", "# - 1) groupby([\"frame_idx\",\"model_name\"]) -> 평균(width,height,area,score), count(*) => df_frame_model\n", "# - 2) 그 df를 이용해 (가로, 세로, area, score, count) 각 지표별 그래프\n", "# - 3) 한꺼번에 그리는 버전(서브플롯)도 저장\n", "\n", "# %%\n", "if df_det.empty:\n", " print(\"⚠️ df_det가 비어있습니다. worker 검출이 없었던 듯합니다.\")\n", "else:\n", " # 1) groupby\n", " df_frame_model = df_det.groupby([\"frame_idx\",\"model_name\"]).agg({\n", " \"width\":\"mean\",\n", " \"height\":\"mean\",\n", " \"area\":\"mean\",\n", " \"score\":\"mean\",\n", " }).rename(columns={\n", " \"width\":\"avg_width\",\n", " \"height\":\"avg_height\",\n", " \"area\":\"avg_area\",\n", " \"score\":\"avg_score\"\n", " }).reset_index()\n", " \n", " # count\n", " df_count = df_det.groupby([\"frame_idx\",\"model_name\"]).size().reset_index(name=\"count\")\n", " df_frame_model = pd.merge(df_frame_model, df_count, on=[\"frame_idx\",\"model_name\"], how=\"left\")\n", " \n", " print(\"=== df_frame_model (프레임 x 모델) ===\")\n", " print(df_frame_model.head(10))\n", " \n", " # 2) 지표별 그래프: frame_idx x-axis\n", " # 모델별 다른 색깔의 라인\n", " metrics = [\"avg_width\",\"avg_height\",\"avg_area\",\"avg_score\",\"count\"]\n", " \n", " # (a) 지표별 그래프 각각 저장\n", " for metric in metrics:\n", " plt.figure(figsize=(8,6))\n", " for m_name in df_frame_model[\"model_name\"].unique():\n", " sub_df = df_frame_model[df_frame_model[\"model_name\"]==m_name].sort_values(\"frame_idx\")\n", " plt.plot(sub_df[\"frame_idx\"], sub_df[metric], marker='o', label=m_name)\n", " \n", " plt.title(f\"{metric} vs frame_idx\")\n", " plt.xlabel(\"Frame_idx\")\n", " plt.ylabel(metric)\n", " plt.legend()\n", " out_fn = os.path.join(output_folder, f\"graph_{metric}_vs_frame.png\")\n", " plt.savefig(out_fn, dpi=150)\n", " plt.show()\n", " \n", " # (b) 한꺼번에 (서브플롯 5개) => 가독성이 떨어질 수 있어도 참고용\n", " fig, axes = plt.subplots(3,2, figsize=(12,12)) # 3행2열 => 6칸중 5칸 사용\n", " ax_list = axes.ravel()\n", " \n", " for i, metric in enumerate(metrics):\n", " ax = ax_list[i]\n", " for m_name in df_frame_model[\"model_name\"].unique():\n", " sub_df = df_frame_model[df_frame_model[\"model_name\"]==m_name].sort_values(\"frame_idx\")\n", " ax.plot(sub_df[\"frame_idx\"], sub_df[metric], marker='o', label=m_name)\n", " ax.set_title(f\"{metric} vs Frame\")\n", " ax.set_xlabel(\"Frame\")\n", " ax.set_ylabel(metric)\n", " ax.legend()\n", " \n", " plt.tight_layout()\n", " sp_fn = os.path.join(output_folder, \"all_metrics_subplots.png\")\n", " plt.savefig(sp_fn, dpi=150)\n", " plt.show()\n", " print(f\"✅ 서브플롯 그래프 저장: {sp_fn}\")\n", " \n", " # CSV 저장\n", " csv_path = os.path.join(output_folder, \"frame_model_stats.csv\")\n", " df_frame_model.to_csv(csv_path, index=False)\n", " print(f\"✅ 프레임별 모델별 통계 CSV 저장: {csv_path}\")\n", " \n", "print(\"🎉 [6] 그래프 & 표 작성 완료.\")\n" ] } ], "metadata": { "kernelspec": { "display_name": "ltb", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.8.20" } }, "nbformat": 4, "nbformat_minor": 2 }