🆕 [2024-10-11] Made Public

FusionSense

[Page] | [Paper] | [Video]

This is the official implementation of FusionSense: Bridging Common Sense, Vision, and Touch for Robust Sparse-View Reconstruction

Irving Fang, Kairui Shi, Xujin He, Siqi Tan, Yifan Wang, Hanwen Zhao, Hung-Jui Huang, Wenzhen Yuan, Chen Feng, Jing Zhang

FusionSense is a novel 3D reconstruction framework that enables robots to fuse priors from foundation models with highly sparse observations from vision and tactile sensors. It enables visually and geometrically accurate scene and object reconstruction, even for conventionally challenging objects.

Installation

Step 1: Install dependencies and Nerfstudio

git clone --recursive https://github.com/ai4ce/FusionSense.git
cd FusionSense
conda env create -f config.yml
conda activate fusionsense

Install compatible pytorch and cuda-toolkit version:

pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit

Install tinycudann:

pip install ninja git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

Step 2: Build Fusionsense

pip install -e .

Step 3: Install and Run Grounded SAM (Need to switch virtual env)

We use Grounded-SAM2 for segmenting the foreground and background. For each

cd Grounded-SAM2-for-masking
cd checkpoints
bash download_ckpts.sh
cd ../gdino_checkpoints
bash download_ckpts.sh

We recommend starting a separate Conda environment for Grounded-SAM2, since Grounded-SAM2 requires cuda=12.1

conda create -n G-SAM-2
conda activate G-SAM-2
conda install pip 
conda install opencv supervision transformers
pip install torch torchvision torchaudio
# select cuda version 12.1
export CUDA_HOME=/path/to/cuda-12.1/
# install Segment Anything 2
pip install -e . 
# install Grounding DINO
pip install --no-build-isolation -e grounding_dino

For further installation problems:

• For dn-splatter, see Installation

• For Grounded-SAM2-for-masking, see Installation

Usage

Select Frames

set train.txt with images id.

Extract Mask

Switch your conda env first
set your scene path and prompt text with the end of '.'
eg. 'transparent white statue.'

conda activate G-SAM-2
cd Grounded-SAM2-for-masking
python grounded_sam2_hf_model_imgs_MaskExtract.py  --path {ABSOLUTE_PATH} --text {TEXT_PROMPT_FOR_TARGET_OBJ}
cd ..

run the script to extract masks.

If the num_no_detection is not 0, you need to select the frame again. Then you will see mask_imgs in /masks, and you can check /annotated frames to see the results more directly.

Run pipeline

conda activate fusionsense
python Initial_Recon.py --data_name {DATASET_NAME} --model_name {MODEL_NAME}

Render outputs

For render jpeg or mp4 outputs using nerfstudio, we recommend install ffmpeg in conda environment:

conda install -c conda-forge x264=='1!161.3030' ffmpeg=4.3.2

To render outputs of pretrained models:

python render_video.py camera-path --load_config your-model-config --camera_path_filename camera_path.json --rendered_output_names rgb depth normal

more details in nerfstudio ns-render.

Dataset Format

datasets/
    ds_name/
    │
    ├── transforms.json # need for training
    │
    ├── train.txt
    │
    ├── images/
    │   ├── rgb_1.png
    │   └── rgb_2.png
    │ 
    ├── realsense_depth/
    │   ├── depth_1.png
    │   └── depth_2.png
    │
    │── tactile/
    │   ├── image
    │   ├── mask
    │   ├── normal
    │   └── patch
    │
    ├── model.stl       # need for evaluation
    │
    ├── normals_from_pretrain/ # generated
    │   ├── rgb_1.png
    │   └── rgb_2.png
    │
    ├── foreground_pcd.ply
    │
    └── merged_pcd.ply

Outputs Format

outputs/
    ds_name/
    │
    ├── MESH/
    │   └── mesh.ply
    │
    ├── nerfstudio_models/
    │   └── 30000.ckpt
    │   
    ├── cluster_centers.npy
    │
    ├── config.yml
    │
    ├── high_grad_pts.pcd
    │
    ├── high_grad_pts_ascii.pcd
    │
    └── dataparser_transforms.json

eval/
    ds_name/ *evaluation results files*