[IROS24] Offical Code for "FruitNeRF: A Unified Neural Radiance Field based Fruit Counting Framework" - Inegrated into Nerfstudio
GPL-3.0 License
Follow these instructions up to and including " tinycudann" to install dependencies and create an environment.
Important: In Section Install nerfstudio please install version 0.3.2 via pip install nerfstudio==0.3.2
not
the latest one!
git clone https://github.com/meyerls/FruitNeRF.git
Navigate to this folder and run python -m pip install -e .
ns-install-cli
Run ns-train -h
: you should see a list of "subcommand" with fruit_nerf included among them.
Please install Grounding-SAM into the segmentation folder. More details can be found in install segment anything and install GroundingDINO. A copied variant is listed below.
# Start from FruitNerf root folder.
cd segmentation
# Clone GroundedSAM repository and rename folder
git clone https://github.com/IDEA-Research/Grounded-Segment-Anything.git grounded_sam
cd grounded_sam
# Checkout version compatible with FruitNeRF
git checkout fe24
You should set the environment variable manually as follows if you want to build a local GPU environment for Grounded-SAM:
export AM_I_DOCKER=False
export BUILD_WITH_CUDA=True
export CUDA_HOME=/path/to/cuda-11.3/
Install Segment Anything:
python -m pip install -e segment_anything
Install Grounding DINO:
pip install --no-build-isolation -e GroundingDINO
Install diffusers and misc:
pip install --upgrade diffusers[torch]
pip install opencv-python pycocotools matplotlib onnxruntime onnx ipykernel
Download pretrained weights
cd .. # Download into grounded_sam
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
wget https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth
Install SAM-HQ
pip install segment-anything-hq
Download SAM-HQ checkpoint from here (We recommend ViT-H HQ-SAM) into the Grounded-Segment-Anything folder.
Done!
Now that FruitNeRF is installed you can start counting fruits! You can use your own data, our real or synthetic FruitNeRF Dataset or the Fuji Dataset. If you use ower FruitNeRF dataset you can skip the preparations step and jump to Training.
For our data and the Fuji dataset you first have to compute the intrinsic and extrinsic camera parameters and segment the images using grounded-SAM:
ns-process-fruit-data --data {path/to/image-dir} --output-dir {path/to/output-dir} --segmentation-class [Str+Str+Str]
--data [PATH]
: Path the data, either a video file or a directory of images.--output-dir [PATH]
: Path to the output directory.--segmentation-class [Str+Str+Str+...]]
Text prompt for segmentation with Grounded SAM. Multiple arguments are--num_downscales [INT]
: Number of times to downscale the images. Default is 3.--text_threshold [FLOAT]
Threshold for text prompt/class to segment images. Default value is 0.15.--box_threshold [FLOAT]
Threshold for bounding box prediction. Default value is 0.15.--data_semantic [PATH]
: Predefined path to precomputed masks.--skip-colmap
: skips COLMAP and generates transforms.json if possible.--skip_image_processing
: skips copying and downscaling of images and only runs COLMAP if possible and enabled.--flag_segmentation_image_debug
: saves the masks overlay on rgb images.If you already have binary segmentation masks please parse the image folder:
ns-prepocess-fruit-data --data {path/to/image-dir} --output-dir {path/to/output-dir} --data_semantic {path/to/seg-dir}
ns-train fruit_nerf --data {path/to/workspace-dir} --output-dir {path/to/output-dir}
ns-train fruit_nerf_big --data {path/to/workspace-dir} --output-dir {path/to/output-dir}
ns-export-semantics semantic-pointcloud --load-config {path/to/config.yaml} --output-dir {path/to/export/dir} --use-bounding-box True --bounding-box-min -1 - 1 -1 --bounding-box-max 1 1 1 --num_rays_per_batch 2000 --num_points_per_side 2000
Clustering is not integrated into the nerfstudio pipeline. Therefore, we have created a specific cluster
script (clustering\run_clustering.py
).
If you want to use it for your own data you have to create a config profile first:
Apple_GT_1024x1024_300 = {
"path": "/path/2/extracted/pcd/semantic_colormap.ply",
"remove_outliers_nb_points": 200, # Clean pcd
"remove_outliers_radius": 0.01, # Clean pcd
"down_sample": 0.001, # Voxel downsample for faster computation / clustering
"eps": 0.01,
"cluster_merge_distance": 0.04, # Merge distance for small clusters
"minimum_size_factor": 0.3,
"min_samples": 100, # Min cluster point size
'template_path': './clustering/apple_template.ply', # Template apple /fruit
'apple_template_size': 0.7, # Scale apple template if no gt size is available
'gt_cluster': "/path/2/gt/mesh/fruits.obj", # or None
"gt_count": 283 # or None
}
Afterward perform the Clustering (see more information in clustering\run_clustering.py
!):
Baum = Apple_GT_1024x1024_300
clustering = Clustering(remove_outliers_nb_points=Baum['remove_outliers_nb_points'],
remove_outliers_radius=Baum['remove_outliers_radius'],
voxel_size_down_sample=Baum['down_sample'],
template_path=Baum['template_path'],
min_samples=Baum['min_samples'],
apple_template_size=Baum['apple_template_size'],
gt_cluster=Baum['gt_cluster'],
cluster_merge_distance=Baum['cluster_merge_distance'],
gt_count=Baum['gt_count']
)
count = clustering.count(pcd=Baum["path"], eps=Baum['eps'])
For reproducibility, we provide the extracted point clouds for our synthetic and real-world data. From Table I and Fig.8. Data can be downloaded from here.
To reproduce our counting results you can download the extracted point clouds for every training run. Download can be found here: tbd.
Link:
Link:
If you find this useful, please cite the paper!