StackCounting Dataset: 3D Stacked Objects with Ground-Truth Count and Geometry

StackCounting Dataset This dataset is part of the Counting Stacked Objects project ( Project page | Paper ), presented at ICCV25 (Oral). Most importantly, it includes multi-view images of stacked objects, as well as ground-truth counts. It includes both large-scale synthetic data for training as well as real data for evaluation. Specifically, this repository contains the following 3 datasets: Stacks-3D-Real: Multi-view images of stacks with ground-truth counts verified manually (45 scenes) Segmentation masks of objects and their containers Human estimations for over 30 participants Stacks-3D-Synth: A large-scale synthetic dataset of over 13000 3D stacks Ground-truth counts and percentage of volume occupied by objects Multi-View 2D renders and masks 3D meshes of the object and stack (Note: some stack meshes are not included due to size limitations, and the rest need to be decompressed with the python script included in this repository) Stacks-2D-Real: A single-view dataset of real-scenes and ground-truth counts (71 scenes) 2D localizations and count of visible objects, i.e. only objects that appear in the image Citation If you find it useful, 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}} This 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},}