| # Data Preparation | |
| ## Preprocessing | |
| The directory should be orgainized as | |
| ``` | |
| Video-3D-LLM # project root | |
| ├── data | |
| │ ├── scannet | |
| │ │ ├── scans | |
| │ │ ├── posed_images | |
| │ │ ├── pcd_with_object_aabbs | |
| │ │ └── mask | |
| │ ├── embodiedscan | |
| │ │ ├── embodiedscan_infos_train.pkl | |
| │ │ ├── embodiedscan_infos_val.pkl | |
| │ │ └── embodiedscan_infos_test.pkl | |
| │ ├── metadata | |
| │ │ ├── scannet_select_frames.json | |
| │ │ ├── pcd_discrete_0.1.pkl | |
| │ │ ├── scannet_train_gt_box.json | |
| │ │ └── scannet_val_pred_box.json | |
| │ ├── prcoessed | |
| │ │ ├── multi3drefer_train_llava_style.json | |
| │ │ ├── multi3drefer_val_llava_style.json | |
| │ │ ├── ... | |
| ``` | |
| ### ScanNet v2 | |
| 1. Download the ScanNet v2 dataset [here](http://www.scan-net.org/). The folder of ScanNet should look like | |
| ``` | |
| Video-3D-LLM # project root | |
| ├── data | |
| │ ├── scannet | |
| │ │ ├── scans | |
| │ │ │ ├── [scene_id] | |
| │ │ │ │ ├── [scene_id]_vh_clean_2.ply | |
| │ │ │ │ ├── [scene_id]_vh_clean_2.0.010000.segs.json | |
| │ │ │ │ ├── [scene_id].aggregation.json | |
| │ │ │ │ ├── [scene_id].txt | |
| │ │ │ │ └── [scene_id].sens | |
| ``` | |
| 2. Extract color images, depth images and camera parameter using the following script, which is modified from [EmbodiedScan](https://github.com/OpenRobotLab/EmbodiedScan/blob/main/embodiedscan/converter/generate_image_scannet.py). | |
| ```bash | |
| python scripts/3d/preprocessing/generate_image_scannet.py --fast | |
| ``` | |
| Please extract the processed color images through `cat posed_images_part** > posed_images.tar.gz`. | |
| 3. Extract point clouds for each scene. | |
| ```bash | |
| python scripts/3d/preprocessing/extract_scannet_pcd.py | |
| ``` | |
| This will generate the point clouds and object bounding boxes for each scan. | |
| ### EmbodiedScan | |
| Download EmbodiedScan data at this [link](https://github.com/OpenRobotLab/EmbodiedScan/tree/main/data). You need to fill out the [official form](https://docs.google.com/forms/d/e/1FAIpQLScUXEDTksGiqHZp31j7Zp7zlCNV7p_08uViwP_Nbzfn3g6hhw/viewform) to get the access to the dataset. Decompress the embodiedscan and the directory should be orgainized as | |
| ``` | |
| ├── data | |
| │ ├── metadata | |
| │ │ ├── embodiedscan | |
| │ │ │ ├── embodiedscan_infos_train.pkl | |
| │ │ │ ├── embodiedscan_infos_val.pkl | |
| │ │ │ └── embodiedscan_infos_test.pkl | |
| ``` | |
| ### Meta Information | |
| 1. Prepare the object proposals. For training set, we directly use the ground truth via the following command. | |
| ```bash | |
| python scripts/3d/preprocessing/extract_gt_box.py | |
| ``` | |
| For the validation set, we utilize the object proposals detected by Mask3D. LEO provided the corresponding annotation results [here](https://huggingface.co/datasets/huangjy-pku/LEO_data/blob/main/mask.zip). We place it at `data/scannet/mask` and process it using the following script. | |
| ```bash | |
| python scripts/3d/preprocessing/extract_pred_box.py | |
| ``` | |
| 2. Prepare the maximum coverage sampling. Firstly we need to preprocess the voxel for each scan for maximum coverage sampling. The results will be saved at `data/metadata/pcd_discrete_0.1.pkl`. | |
| ```bash | |
| python scripts/3d/preprocessing/convert_pcd_to_voxel.py | |
| ``` | |
| And then we perform the maximum coverage sampling offiline, and the results will be saved at `data/metadata/scannet_select_frames.json`. | |
| ``` | |
| python scripts/3d/preprocessing/max_coverage_sampling.py | |
| ``` | |
| ### Downstream Benchmarks | |
| 1. SQA3D: Download the [SQA3D](https://github.com/SilongYong/SQA3D?tab=readme-ov-file) and convert the annotation to the LLaVA format using the following script. | |
| ```bash | |
| python scripts/3d/preprocessing/process_sqa3d.py | |
| ``` | |
| 2. ScanQA: Download the [ScanQA](https://github.com/ATR-DBI/ScanQA/blob/main/docs/dataset.md) and convert the annotation using the following script. | |
| ```bash | |
| python scripts/3d/preprocessing/process_scanqa.py | |
| ``` | |
| 3. ScanRefer: Download the [ScanRefer](https://daveredrum.github.io/ScanRefer/), and then run the following command. | |
| ```bash | |
| python scripts/3d/preprocessing/process_scanrefer.py | |
| ``` | |
| 4. Scan2Cap: Convert the annotation of ScanRefer to Scan2Cap. | |
| ```bash | |
| python scripts/3d/preprocessing/process_scan2cap.py | |
| ``` | |
| 5. Multi3DRefer: Download the [Multi3DRefer](https://github.com/3dlg-hcvc/M3DRef-CLIP). | |
| ```bash | |
| python scripts/3d/preprocessing/process_multi3drefer.py | |
| ``` |