Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations
This repository contains the model weights for Concerto, a novel approach for learning robust spatial representations presented in the paper Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations.
- Paper: Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations
- Project Page: https://pointcept.github.io/Concerto/
- Codebase: https://github.com/Pointcept/Pointcept
- Inference: https://github.com/Pointcept/Concerto
Models
concerto_base_siglip.pth: replace DINOv2 with SigLIPv2 in cross-modal distillation.
concerto_large_linear_prob_head_sc.pth: ScanNet linear probing head.
concerto_large_outdoor.pth: outdoor pretraining weight using similar recipe.
The default models(concerto_large/base/small/tiny) are the pre-release version of our next work, Utonia, which can deal with input without color and normal. If colors and normals are not available, please set them to zeros. We pre-release these for general public use because many tasks lack such information. The original Concerto model is concerto_base_origin.pth.
| Model Size | Channels | Depths | Head nums |
|---|---|---|---|
| Tiny | (16, 32, 64, 128, 256) | (1, 1, 1, 3, 1) | (1, 2, 4, 8, 16) |
| Small | (32, 64, 128, 256, 512) | (2, 2, 2, 6, 2) | (2, 4, 8, 16, 32) |
| Base | (48, 96, 192, 384, 512) | (3, 3, 3, 12, 3) | (3, 6, 12, 24, 32) |
| Large | (64, 128, 256, 512, 768) | (3, 3, 3, 12, 3) | (4, 8, 16, 32, 48) |
Abstract
Humans learn abstract concepts through multisensory synergy, and once formed, such representations can often be recalled from a single modality. Inspired by this principle, we introduce Concerto, a minimalist simulation of human concept learning for spatial cognition, combining 3D intra-modal self-distillation with 2D-3D cross-modal joint embedding. Despite its simplicity, Concerto learns more coherent and informative spatial features, as demonstrated by zero-shot visualizations. It outperforms both standalone SOTA 2D and 3D self-supervised models by 14.2% and 4.8%, respectively, as well as their feature concatenation, in linear probing for 3D scene perception. With full fine-tuning, Concerto sets new SOTA results across multiple scene understanding benchmarks (e.g., 80.7% mIoU on ScanNet). We further present a variant of Concerto tailored for video-lifted point cloud spatial understanding, and a translator that linearly projects Concerto representations into CLIP's language space, enabling open-world perception. These results highlight that Concerto emerges spatial representations with superior fine-grained geometric and semantic consistency.
Usage
For detailed installation, data preparation, training, and testing instructions, please refer to the official codebase and inference demo.
Citation
If you find Concerto or the Pointcept codebase useful in your research, please cite the following papers:
@misc{pointcept2023,
title={Pointcept: A Codebase for Point Cloud Perception Research},
author={Pointcept Contributors},
howpublished = {\url{https://github.com/Pointcept/Pointcept}},
year={2023}
}
@article{zhang2025concerto,
title={Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations},
author={Zhang, Yujia and Wu, Xiaoyang and Lao, Yixing and Wang, Chengyao and Tian, Zhuotao and Wang, Naiyan and Zhao, Hengshuang},
journal={Conference on Neural Information Processing Systems},
year={2025},
}