Question about training on original DL3DV dataset with AnySplat

[Frank-Wang-wow]

Hi,

We are trying to run AnySplat with the original DL3DV dataset.
We have downloaded the full dataset from the official source, and the structure we have is:

dl3dv/
├── 1k/
│ ├── transforms.json
│ ├── frame_0001.jpg
│ ├── frame_0002.jpg
│ └── ...
├── 2k/
│ └── ...
├── ...
└── 11k/
└── ...

However, in the AnySplat code, the DatasetDL3DV class seems to expect a different directory structure, including train_index.json, test_index.json, images/, images_4/, and depth/ folders inside each scene directory.

With the dataset we downloaded, we cannot run the code directly because:

• There are no train_index.json / test_index.json files.
• Images are stored directly in the scene folder (e.g., frame_0001.jpg) instead of inside images/ or images_4/.
• Depth .npy files are missing.

Could you please clarify:

1. How did you process the original DL3DV dataset to match the structure expected by AnySplat?
2. Is there any preprocessing script you used to generate train_index.json / test_index.json and reorganize the images/depth files?
3. If possible, could you share the code or instructions for this preprocessing step?

Thank you!

[dev111ce]

May I ask how you finally solved this problem?

[dev111ce]

You can generate train_index.json/test_index.json like this.
`
import os
import json

数据集根目录（根据你的实际路径修改）

data_root = "datasets/DL3DV-10K/1K_480"

获取所有场景文件夹

scenes = [f for f in os.listdir(data_root) if os.path.isdir(os.path.join(data_root, f))]
scenes.sort() # 按名称排序

划分训练集和测试集（例如前90%为训练，后10%为测试）

split_idx = int(len(scenes) * 0.7)
train_scenes = scenes[:split_idx]
test_scenes = scenes[split_idx:]

生成train_index.json

with open(os.path.join(data_root, "train_index.json"), "w") as f:
json.dump(train_scenes, f, indent=2)

生成test_index.json

with open(os.path.join(data_root, "test_index.json"), "w") as f:
json.dump(test_scenes, f, indent=2)

print(f"生成完成：{len(train_scenes)}个训练场景，{len(test_scenes)}个测试场景")
`