{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [Code cell 0]: Imports & paths\n", "import os\n", "import sys\n", "import shutil\n", "import subprocess\n", "import numpy as np\n", "import cv2\n", "from PIL import Image\n", "import matplotlib.pyplot as plt\n", "import torch\n", "\n", "# Distill-Any-Depth 관련 경로\n", "DISTILL_BASE_DIR = \"/ltb/media/ltb/90887173887158A4/Users/ltb/모델들/Distill-Any-Depth\"\n", "DISTILL_INFER_SCRIPT = os.path.join(DISTILL_BASE_DIR, \"scripts/00_infer.sh\")\n", "DISTILL_INPUT_DIR = os.path.join(DISTILL_BASE_DIR, \"data\", \"input\")\n", "DISTILL_OUTPUT_DIR = os.path.join(DISTILL_BASE_DIR, \"output\", \"large\")\n", "\n", "# Depth-Anything-V2 관련 경로\n", "DEPTH_ANYTHING_DIR = \"/ltb/media/ltb/90887173887158A4/Users/ltb/회사용/4090_옮길거/workspace/Depth-Anything-V2\"\n", "DEPTH_ANYTHING_CHECKPOINT_VITL = os.path.join(DEPTH_ANYTHING_DIR, \"checkpoints/depth_anything_v2_vitl.pth\")\n", "DEPTH_ENCODER_TYPE = \"vitl\"\n", "\n", "# 입력 이미지 폴더 (Distill-Any-Depth와 동일 폴더로 가정)\n", "INPUT_FOLDER = DISTILL_INPUT_DIR # \"/.../Distill-Any-Depth/data/input\"\n", "# 최종 비교 결과 저장 폴더\n", "COMPARE_DEPTH_SAVE_DIR = \"/ltb/media/ltb/90887173887158A4/Users/ltb/compare_depth\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [1] Distill-Any-Depth 추론\n", "def run_distill_any_depth():\n", " print(\"[Distill-Any-Depth] 전체 이미지 추론을 시작합니다.\")\n", " if os.path.exists(DISTILL_OUTPUT_DIR):\n", " for old_file in os.listdir(DISTILL_OUTPUT_DIR):\n", " path_to_remove = os.path.join(DISTILL_OUTPUT_DIR, old_file)\n", " if os.path.isfile(path_to_remove):\n", " os.remove(path_to_remove)\n", " else:\n", " shutil.rmtree(path_to_remove, ignore_errors=True)\n", " else:\n", " os.makedirs(DISTILL_OUTPUT_DIR, exist_ok=True)\n", "\n", " # 실행\n", " prev_cwd = os.getcwd()\n", " os.chdir(DISTILL_BASE_DIR)\n", " try:\n", " cmd = f\"bash {DISTILL_INFER_SCRIPT}\"\n", " subprocess.run(cmd, shell=True, check=True)\n", " except subprocess.CalledProcessError as e:\n", " print(\"[Distill-Any-Depth] 추론 에러 발생:\", e)\n", " finally:\n", " os.chdir(prev_cwd)\n", "\n", " print(\"[Distill-Any-Depth] 추론 완료!\")\n", "\n", " # image_logs -> compare_depth 복사\n", " distill_image_logs = os.path.join(DISTILL_OUTPUT_DIR, \"image_logs\")\n", " if not os.path.exists(distill_image_logs):\n", " print(f\"[Distill-Any-Depth] image_logs 폴더가 없습니다: {distill_image_logs}\")\n", " else:\n", " os.makedirs(COMPARE_DEPTH_SAVE_DIR, exist_ok=True)\n", " distill_logs = os.listdir(distill_image_logs)\n", " for f in distill_logs:\n", " src_path = os.path.join(distill_image_logs, f)\n", " if os.path.isfile(src_path):\n", " dst_name = f\"distill_{f}\"\n", " dst_path = os.path.join(COMPARE_DEPTH_SAVE_DIR, dst_name)\n", " shutil.copy2(src_path, dst_path)\n", " print(f\"[Distill-Any-Depth] Copied => {os.path.abspath(dst_path)}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [2] Load Depth-Anything-V2\n", "def load_depth_anything_v2(checkpoint_path, device=\"cuda\"):\n", " if DEPTH_ANYTHING_DIR not in sys.path:\n", " sys.path.append(DEPTH_ANYTHING_DIR)\n", "\n", " from depth_anything_v2.dpt import DepthAnythingV2\n", " model = DepthAnythingV2(\n", " encoder=DEPTH_ENCODER_TYPE,\n", " features=256,\n", " out_channels=[256, 512, 1024, 1024]\n", " )\n", " model.load_state_dict(torch.load(checkpoint_path, map_location=device))\n", " model = model.to(device)\n", " model.eval()\n", " return model" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [3] V2 inference on input images\n", "def run_depth_anything_on_images(model, input_images, device=\"cuda\", resize_size=700):\n", " depth_results = {}\n", " with torch.no_grad():\n", " for img_path in input_images:\n", " img_name = os.path.basename(img_path)\n", " pil_img = Image.open(img_path).convert('RGB')\n", " img_np = np.array(pil_img)\n", " depth_map = model.infer_image(img_np, resize_size)\n", " depth_map = depth_map.squeeze()\n", " depth_results[img_name] = depth_map\n", " return depth_results\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [4] Util\n", "def normalize_minmax(depth_map):\n", " d_min, d_max = depth_map.min(), depth_map.max()\n", " if d_max - d_min < 1e-6:\n", " return np.zeros_like(depth_map, dtype=np.float32)\n", " return (depth_map - d_min) / (d_max - d_min)\n", "\n", "def depth_to_colormap(depth_map, cmap_name=\"plasma\"):\n", " depth_map_norm = normalize_minmax(depth_map)\n", " cmap = plt.get_cmap(cmap_name)\n", " depth_color = cmap(depth_map_norm)[:, :, :3]\n", " depth_color = (depth_color * 255).astype(np.uint8)\n", " return depth_color\n", "\n", "def create_comparison_image(original_img, depth_img_1, depth_img_2, label1=\"ModelA\", label2=\"ModelB\"):\n", " h, w, _ = original_img.shape\n", " concat_img = np.concatenate([original_img, depth_img_1, depth_img_2], axis=1)\n", "\n", " import cv2\n", " font = cv2.FONT_HERSHEY_SIMPLEX\n", " font_scale = 1.0\n", " color_text = (255, 255, 255)\n", " thickness = 2\n", "\n", " cv2.putText(concat_img, \"Original\", (10, 30), font, font_scale, color_text, thickness, cv2.LINE_AA)\n", " label_x = w + 10\n", " cv2.putText(concat_img, label1, (label_x, 30), font, font_scale, color_text, thickness, cv2.LINE_AA)\n", " label_x2 = w*2 + 10\n", " cv2.putText(concat_img, label2, (label_x2, 30), font, font_scale, color_text, thickness, cv2.LINE_AA)\n", " return concat_img\n", "\n", "def create_difference_map(depth_v2_color, distill_color):\n", " diff_img = cv2.absdiff(depth_v2_color, distill_color)\n", " return diff_img" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [5] 수치 기반 + RGB 기반 비교\n", "def run_depth_comparison_on_input_folder():\n", " device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n", "\n", " # 1) Distill-Any-Depth 추론\n", " run_distill_any_depth()\n", "\n", " # 2) Load Depth-Anything-V2\n", " model_v2 = load_depth_anything_v2(checkpoint_path=DEPTH_ANYTHING_CHECKPOINT_VITL, device=device)\n", "\n", " os.makedirs(COMPARE_DEPTH_SAVE_DIR, exist_ok=True)\n", "\n", " input_files = [f for f in os.listdir(INPUT_FOLDER)\n", " if f.lower().endswith(('.jpg','.jpeg','.png','.bmp','.webp'))]\n", " input_files.sort()\n", " full_paths = [os.path.join(INPUT_FOLDER, x) for x in input_files]\n", "\n", " # 3) V2 inference\n", " depth_results_v2 = run_depth_anything_on_images(model_v2, full_paths, device=device, resize_size=700)\n", "\n", " # distill logs\n", " distill_image_logs = os.path.join(DISTILL_OUTPUT_DIR, \"image_logs\")\n", " if not os.path.exists(distill_image_logs):\n", " print(f\"[경고] {distill_image_logs} 폴더가 없어 Distill 결과 없음.\")\n", " return\n", "\n", " distill_outputs_jpg = sorted([f for f in os.listdir(distill_image_logs) if f.startswith(\"da_sota_\") and f.endswith(\".jpg\")])\n", " distill_outputs_npy = sorted([f for f in os.listdir(distill_image_logs) if f.startswith(\"da_sota_\") and f.endswith(\"_depth.npy\")])\n", "\n", " for idx, in_file in enumerate(input_files):\n", " in_path = os.path.join(INPUT_FOLDER, in_file)\n", " origin_pil = Image.open(in_path).convert(\"RGB\")\n", " origin_np = np.array(origin_pil)\n", " h, w, c = origin_np.shape\n", "\n", " # V2 depth\n", " depth_map_v2 = depth_results_v2.get(in_file, None)\n", " if depth_map_v2 is None:\n", " print(f\"[오류] V2 depth 결과 없음: {in_file}\")\n", " continue\n", " # v2 컬러맵\n", " depth_color_v2 = depth_to_colormap(depth_map_v2, \"plasma\")\n", " depth_color_v2 = cv2.resize(depth_color_v2, (w, h), interpolation=cv2.INTER_LINEAR)\n", "\n", " # Distill JPG\n", " if idx >= len(distill_outputs_jpg):\n", " print(f\"[오류] Distill JPG 인덱스 초과: idx={idx}\")\n", " continue\n", " distill_file_jpg = distill_outputs_jpg[idx]\n", " distill_img_path = os.path.join(distill_image_logs, distill_file_jpg)\n", " if not os.path.isfile(distill_img_path):\n", " print(f\"[오류] Distill JPG 없음: {distill_img_path}\")\n", " continue\n", "\n", " distill_pil = Image.open(distill_img_path).convert(\"RGB\")\n", " distill_np = np.array(distill_pil)\n", " distill_np = cv2.resize(distill_np, (w, h), interpolation=cv2.INTER_LINEAR)\n", "\n", " # Distill NPY\n", " if idx >= len(distill_outputs_npy):\n", " print(f\"[오류] Distill NPY 인덱스 초과 idx={idx}\")\n", " continue\n", " distill_file_npy = distill_outputs_npy[idx]\n", " distill_npy_path = os.path.join(distill_image_logs, distill_file_npy)\n", " if not os.path.isfile(distill_npy_path):\n", " print(f\"[오류] Distill NPY 없음: {distill_npy_path}\")\n", " continue\n", "\n", " distill_depth = np.load(distill_npy_path)\n", " # shape 검사\n", " if distill_depth.ndim == 3 and distill_depth.shape[0] == 1:\n", " distill_depth = distill_depth[0]\n", " dh, dw = distill_depth.shape\n", " if dh < 1 or dw < 1:\n", " print(f\"[오류] Distill depth 크기 이상: shape={distill_depth.shape}\")\n", " continue\n", "\n", " if (dh, dw) != (h, w):\n", " distill_depth = cv2.resize(distill_depth, (w, h), interpolation=cv2.INTER_LINEAR)\n", "\n", " # V2도 0~1로 가정 => 필요시 수동 정규화\n", " v2_min, v2_max = depth_map_v2.min(), depth_map_v2.max()\n", " if v2_max - v2_min < 1e-9:\n", " v2_norm = np.zeros_like(depth_map_v2, dtype=np.float32)\n", " else:\n", " v2_norm = (depth_map_v2 - v2_min)/(v2_max - v2_min)\n", "\n", " # MAE\n", " diff_map_num = np.abs(v2_norm - distill_depth)\n", " mae = diff_map_num.mean()\n", " print(f\"[{in_file}] => MAE(Distill vs V2)={mae:.4f}\")\n", "\n", " diff_map_255 = (diff_map_num*255).astype(np.uint8)\n", " diff_color = cv2.applyColorMap(diff_map_255, cv2.COLORMAP_JET)\n", "\n", " # RGB 차이\n", " rgb_diff_img = create_difference_map(depth_color_v2, distill_np)\n", " compare_3way = create_comparison_image(origin_np, depth_color_v2, distill_np, \"DepthV2\", \"Distill\")\n", "\n", " out_name_base = os.path.splitext(in_file)[0]\n", " out_v2_path = os.path.join(COMPARE_DEPTH_SAVE_DIR, f\"{out_name_base}_v2.png\")\n", " out_distill_path = os.path.join(COMPARE_DEPTH_SAVE_DIR, f\"{out_name_base}_distill.png\")\n", " out_compare_path = os.path.join(COMPARE_DEPTH_SAVE_DIR, f\"{out_name_base}_compare.png\")\n", " out_rgbdiff_path = os.path.join(COMPARE_DEPTH_SAVE_DIR, f\"{out_name_base}_diff_rgb.png\")\n", " out_numdiff_path = os.path.join(COMPARE_DEPTH_SAVE_DIR, f\"{out_name_base}_diff_numerical.png\")\n", "\n", " cv2.imwrite(out_v2_path, cv2.cvtColor(depth_color_v2, cv2.COLOR_RGB2BGR))\n", " cv2.imwrite(out_distill_path, cv2.cvtColor(distill_np, cv2.COLOR_RGB2BGR))\n", " cv2.imwrite(out_compare_path, cv2.cvtColor(compare_3way, cv2.COLOR_RGB2BGR))\n", " cv2.imwrite(out_rgbdiff_path, cv2.cvtColor(rgb_diff_img, cv2.COLOR_RGB2BGR))\n", " cv2.imwrite(out_numdiff_path, diff_color)\n", "\n", " print(f\"[{in_file}] => 저장 완료(MAE={mae:.4f}) => {out_numdiff_path}\")\n", "\n", " print(\"=== 모든 이미지 비교 완료 ===\")" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# %% [6] 실행\n", "if __name__ == \"__main__\":\n", " run_depth_comparison_on_input_folder()" ] } ], "metadata": { "kernelspec": { "display_name": "sam2", "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 }