[ICCV25 Oral] Counting Stacked Objects

From multi-view images of a stack, count the total number of objects!

Project page | Dataset page | stack-dataset-generation repository

Installation

This repository was tested with torch 2.1.2+cu118 and nerfstudio 1.1.5 on Ubuntu 24.04.

1. Since our volume reconstruction is based on nerfstudio's implementation of 3DGS, you will need a nerfstudio environment. You can find instructions here. Make sure you can run splatfacto before proceeding to the next steps: ns-train splatfacto --data data/pasta

Some fixes for common issues with nerfstudio:

• To install tiny-cuda-nn it may be useful to downgrade g++:

sudo apt install gcc-11 g++-11
export CC=/usr/bin/gcc-11
export CXX=/usr/bin/g++-11

• If you're running this on a machine with limited CPU and the command fails when running splatfacto for the first time, it may help to pre-emptively build CUDA code with pip install git+https://github.com/nerfstudio-project/gsplat.git.

2. Install the counting-splatfacto method, which is simply splatfacto with a couple utilities added like saving accumulation and depth maps.

pip install -e .
ns-install-cli

Check you can run the new method: ns-train counting-splatfacto --data data/pasta

3. Download the density net weights and the weights for depth_anything_v2_vitl.pth and put them both in a weights/ at the root of this repository.

4. Download DinoV2 and DepthAnythingV2 in ext. For example:

mkdir ext
cd ext
git clone https://github.com/facebookresearch/dinov2
cd dinov2
git checkout b48308a394a04ccb9c4dd3a1f0a4daa1ce0579b8 
pip install fvcore omegaconf
cd ..
git clone https://github.com/DepthAnything/Depth-Anything-V2/
mv Depth-Anything-V2 DepthAnythingV2

You can make sure that this step worked by running

python counting_3d/utils/compute_depth.py --image-input data/pasta/images/frame_00001.jpg --image-output test.png

Inference

Run the script: . process_scene.sh data/pasta

You can also download more scenes from our dataset.

This repository brought a few minor changes to the original method, so in some cases, you may obtain results slightly different from the original paper.

Data preparation

Important requirements:

• Each scene must have multi-view images, a reference measurement in centimeters to adjust for scale, and known volume of a single object v. If v is not known, it can be estimated using a separate set of images containing a single object and the same volume estimation used in process_scene.sh.
• This step will require the nerfstudio environment described above, plus:

pip install trimesh hydra scikit-image

mkdir weights
mkdir ext
cd ext
git clone git@github.com:facebookresearch/sam2.git
cd ..

• Download sam2.1_hiera_large.pt and put it in the weights folder

Data preparation:

• Organize your data as follows: path/to/folder/, where folder contains path/to/folder/input1, path/to/folder/input2, etc, and input1 contains all images for this scene.
• Write your path/to/folder/ in script data_preparation/data_prep.sh and run it from the root of this repository: . data_preparation/data_prep.sh
• Important: to scale the cameras, we use a simple heuristic to measure a distance in 3DGS space, and compare it with your reference measurement. When 3DGS is training, use the provided UI to measure the distance in 3DGS that matches your reference measurement:
  • Open the viewer at http://localhost:7007/.
  • Once reconstruction is satisfying, click "Pause Training".
  • Press the Click button, then click at the start of the measurement.
  • Rotate the view, then press Click and click and the end of the measurement. The distance in camera scale is printed in the terminal.
  • Save that number, then interrupt the reconstruction with CTRL+C.
  • Enter this number and the real-world measurement when prompted.
• Select the OBJECTS only with left click. Use right clicks for negative prompts. Then close the window, and the segmentation will be propagated to other frames.
• If the objects are in a container, repeat this step for the container+objects. Segment both of them together, and close the window. Please make sure that the segmentations are satisfying before proceeding.

Training

To be added soon...

Dataset generation

Have a look at our stack-dataset-generation repository.

Citation

If this repository is helpful to you, please consider citing the associated publication:

@inproceedings{dumery2025counting,
   title = {{Counting Stacked Objects}},
   author = {Dumery, Corentin and Ett{\'e}, Noa and Fan, Aoxiang and Li, Ren and Xu, Jingyi and Le, Hieu and Fua, Pascal},
   booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
   year = {2025}
}

The released dataset can also directly be cited with:

@misc{dumery2025stackcounting,
   title = {StackCounting Dataset},
   author = {Dumery, Corentin and Ett{\'e}, Noa and D'Alessandro, Adriano},
   year = {2025}
   publisher = {Zenodo},
   doi = {10.5281/zenodo.15609540},
   url = {https://doi.org/10.5281/zenodo.15609540},
}