Gaussian Wild 运行自己的数据集

Gaussian-wild

文件目录

sacre_coeur/
├── dense/
│ ├── images/
│ ├── sparse/不需要/0/
│ ├── stereo/非必需
│
├──sacre.tsv

需要自己写一个生成tsv文件的代码

import os

def read_cameras_text(cameras_txt_path):
    """
    从 COLMAP 的 cameras.txt 中读取相机信息
    """
    cameras = {}
    if not os.path.exists(cameras_txt_path):
        raise FileNotFoundError(f"{cameras_txt_path} 不存在，请检查路径。")

    with open(cameras_txt_path, "r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            # 跳过注释和空行
            if line.startswith("#") or len(line) == 0:
                continue
            elems = line.split()
            camera_id = int(elems[0])
            model = elems[1]
            width = int(elems[2])
            height = int(elems[3])
            params = list(map(float, elems[4:]))

            cameras[camera_id] = {
                "model": model,
                "width": width,
                "height": height,
                "params": params
            }
    return cameras


def read_images_text(images_txt_path):
    """
    从 COLMAP 的 images.txt 中读取图像信息（位姿等）
    """
    images = []
    if not os.path.exists(images_txt_path):
        raise FileNotFoundError(f"{images_txt_path} 不存在，请检查路径。")

    with open(images_txt_path, "r", encoding="utf-8") as f:
        lines = f.readlines()

    idx = 0
    while idx < len(lines):
        line = lines[idx].strip()
        # 跳过注释和空行
        if line.startswith("#") or len(line) == 0:
            idx += 1
            continue

        # 解析图像的第一行
        elems = line.split()
        image_id = int(elems[0])
        # qw, qx, qy, qz
        q = list(map(float, elems[1:5]))
        # tx, ty, tz
        t = list(map(float, elems[5:8]))
        camera_id = int(elems[8])
        # 剩余部分是图像文件名
        image_name = " ".join(elems[9:])

        images.append({
            "image_id": image_id,
            "q": q,
            "t": t,
            "camera_id": camera_id,
            "image_name": image_name
        })

        idx += 1
        # 跳过下一行 (2D特征点信息)
        if idx < len(lines):
            idx += 1

    return images


def generate_tsv_file(cameras, images, output_path):
    """
    生成符合 GSW 需求的 TSV 文件，包含最关键的三列:
      - filename:  对应原本的 image_name
      - id:        对应原本的 image_id
      - split:     表示训练/测试集 (如果没有区分需求，可一律设为train)
    
    其余信息可按需保留:
      camera_id, width, height, model, params, q, t
    """
    with open(output_path, "w", encoding="utf-8") as fid:
        # 注意列顺序：GSW会用到 filename / id / split
        fid.write("filename\tid\tsplit\tcamera_id\twidth\theight\tmodel\tparams\tq\tt\n")
        for img in images:
            cam = cameras[img["camera_id"]]
            cam_params_str = " ".join(map(str, cam["params"]))
            q_str = " ".join(map(str, img["q"]))
            t_str = " ".join(map(str, img["t"]))

            # 如果不区分 train/test，可全部写成 "train"
            # 或者你也可以根据某种逻辑判断是 "train" 还是 "test"
            split_str = "train"

            row = (
                f"{img['image_name']}\t"    # filename
                f"{img['image_id']}\t"      # id
                f"{split_str}\t"           # split
                f"{img['camera_id']}\t"
                f"{cam['width']}\t"
                f"{cam['height']}\t"
                f"{cam['model']}\t"
                f"{cam_params_str}\t"
                f"{q_str}\t"
                f"{t_str}\n"
            )
            fid.write(row)


if __name__ == "__main__":
    # 假设你的txt文件在 <current_directory>/sparse/ 下
    current_directory = os.path.dirname(os.path.abspath(__file__))
    colmap_data_path = os.path.join(current_directory, "sparse")

    cameras_txt_path = os.path.join(colmap_data_path, "cameras.txt")
    images_txt_path = os.path.join(colmap_data_path, "images.txt")

    # 生成的 TSV，直接放在当前脚本同级目录下
    output_tsv_path = os.path.join(current_directory, "images.tsv")

    # 读取相机与图像信息
    cameras_dict = read_cameras_text(cameras_txt_path)
    images_list = read_images_text(images_txt_path)

    # 生成 TSV (包含 filename, id, split 等列)
    generate_tsv_file(cameras_dict, images_list, output_tsv_path)

    print(f"TSV 文件已生成: {output_tsv_path}")

之后要将tsv文件放到root_dir所在目录的上级目录

之后我们要把自己的datasets伪装成官方的数据集来运行

写了个自动化的bash指令来运行全部程序

#!/bin/bash

# 设置场景名称和路径
SCENE_NAME="sacre_coeur"
DATA_NAME="tum" #在这里修改为自己的数据集
SOURCE_PATH="/root/autodl-fs/Gaussian-Wild/data/${DATA_NAME}/dense"  
MODEL_PATH="outputs/${DATA_NAME}/model"


# 创建输出目录
mkdir -p ${MODEL_PATH}

# 训练
echo "Starting training for ${SCENE_NAME}..."
CUDA_VISIBLE_DEVICES=0 python ./train.py \
    --source_path ${SOURCE_PATH} \
    --scene_name ${SCENE_NAME} \
    --model_path ${MODEL_PATH} \
    --eval \
    --resolution 2 \
    --iterations 10000

# 渲染结果
echo "Rendering training and testing results..."
CUDA_VISIBLE_DEVICES=0 python ./render.py \
    --model_path ${MODEL_PATH}

# # 渲染多视角视频
# echo "Rendering multi-view video..."
# CUDA_VISIBLE_DEVICES=0 python ./render.py \
#     --model_path ${MODEL_PATH} \
#     --skip_train \
#     --skip_test \
#     --render_multiview_vedio

# # 渲染外观调整效果
# echo "Rendering appearance tuning..."
# CUDA_VISIBLE_DEVICES=0 python ./render.py \
#     --model_path ${MODEL_PATH} \
#     --skip_train \
#     --skip_test \
#     --render_interpolate

# # 评估右半部分
# echo "Evaluating right half metrics..."
# CUDA_VISIBLE_DEVICES=0 python ./metrics_half.py \
#     --model_path ${MODEL_PATH}

# # 评估全图像
# echo "Evaluating full image metrics..."
# CUDA_VISIBLE_DEVICES=0 python ./metrics.py \
#     --model_path ${MODEL_PATH}

echo "All tasks completed!"