🛠️ Installation Guide#

😄 Don’t worry — both Quick Installation and Dataset Preparation are beginner-friendly.

Quick Installation#

git clone https://github.com/InternRobotics/InternSR.git
cd InternSR
pip install -e .

Dataset Preparation#

We recommend placing all data under data/. The expected directory structure under data/ is as follows :

data/
├── images/ # `images/` folder stores all image modality files from the datasets
├── videos/ # `videos/` folder contains all video modality files from the datasets
├── annotations/ # `annotations/` folder holds all text annotation files from the datasets

MMScan#

Download the image zip files from Hugging Face (~56G), combine and unzip them under ./data/images/mmscan.
Download the annotations from Hugging Face and place them under ./data/annotations.

data/
├── images/
│   ├── mmscan/
│   │   ├── 3rscan
│   │   ├── 3rscan_depth
│   │   ├── matterport3d
│   │   ├── scannet
├── annotations/
│   ├── embodiedscan_video_meta/
│   ├── ├── image.json
│   ├── ├── depth.json
│   ├── ├── ...
│   ├── mmscan_qa_val_0.1.json
│   ├── ...

Note: The file mmscan_qa_val_{ratio}.json contains the validation data at the specified ratio.

OST-Bench#

Download the images from Hugging Face/Kaggle(~5G) and download the .tsv file , place them as follows:

data/
├── images/
│   ├── OST/
│   │   ├── <scan_id>
│   │   ├── ...
├── annotations/
│   ├── OST.tsv

MMSI-Bench#

Download the .tsv file (~1G, including images) , place it as follows:

data/
├── annotations/
│   ├── MMSI_Bench.tsv

EgoExo-Bench#

Download the processed video data from the Hugging Face.
Due to license restrictions, data from the Ego-Exo4D project is not included. Users should acquire it separately by following the official Ego-Exo4D guidelines.
Download the .tsv file , place them as follows:

data/
├── videos/
│   ├── EgoExo4D/tasks
│   ├── processed_frames
│   ├── processed_video
├── annotations/ EgoExoBench_MCQ.tsv