Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,63 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# DeFM: Learning Foundation Representations from Depth for Robotics
|
| 2 |
+
|
| 3 |
+
<div align="center">
|
| 4 |
+
|
| 5 |
+
[](https://www.python.org/downloads/)
|
| 6 |
+
[](https://pytorch.org/get-started/locally/)
|
| 7 |
+
[](https://github.com/astral-sh/uv)
|
| 8 |
+
[](https://opensource.org/licenses/Apache-2.0)
|
| 9 |
+
[](https://github.com/leggedrobotics/defm)
|
| 10 |
+
[](TODO-link)
|
| 11 |
+
[](https://leggedrobotics.github.io/defm)
|
| 12 |
+
</div>
|
| 13 |
+
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
**DeFM** (Depth Foundation Model) is a vision backbone trained on **60M depth images** via self-distillation. It is engineered for robotic perception, providing metric-aware representations that excel in sim-to-real transfer and cross-sensor generalization.
|
| 17 |
+
|
| 18 |
+
TL;DR - A DINO-style encoder, but for depth image inputs.
|
| 19 |
+
|
| 20 |
+
## 🌟 Key Features
|
| 21 |
+
- **Large-Scale Pretraining**: We pretrain on our curated dataset of 60 M depth images using self-distillation.
|
| 22 |
+
- **Semantic Awareness**: DeFM learns not only robust geometric priors but also semantically rich features from just depth images.
|
| 23 |
+
- **Metric-Aware Normalization**: Our novel three channel input normalization preserves metric depth across multiple scales.
|
| 24 |
+
- **Compact efficient models**: We distill our DeFM-ViT-L into a family of smaller efficient CNNs as small as 3M params for robot policy learning.
|
| 25 |
+
- **Robotics Proven**: Our encoder is proven effective for diverse robotic tasks such as navigation, manipulation and locomotion without task-specific fine-tuning.
|
| 26 |
+
|
| 27 |
+
## Usage
|
| 28 |
+
|
| 29 |
+
Visit our [github repo](https://github.com/leggedrobotics/defm) for details on how to use the models.
|
| 30 |
+
|
| 31 |
+
## 📊 Model Zoo
|
| 32 |
+
|
| 33 |
+
The following table provides a comprehensive overview of the DeFM model family, including architectural parameters, inference latency across training and deployment hardware (224x224), and performance on the ImageNet-1k-Depth benchmark.
|
| 34 |
+
|
| 35 |
+
| Model | Params (M) | RTX 4090 (ms) | Jetson Orin (ms) | Top-5 KNN (%) | Linear Prob (%) | Checkpoint |
|
| 36 |
+
| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
|
| 37 |
+
| **DeFM ViT-L/14** | 307.0 | 624.91 | 72.82 | 84.79 | 71.72 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_vit_l14.pth) |
|
| 38 |
+
| **DeFM ViT-S/14** | 22.1 | 63.76 | 11.92 | 78.06 | 61.54 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_vit_s14.pth) |
|
| 39 |
+
| **DeFM ResNet-50** | 26.2 | 69.39 | 17.79 | 77.63 | 61.54 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_resnet50.pth) |
|
| 40 |
+
| **DeFM ResNet-34** | 21.8 | 33.08 | 13.54 | 72.72 | 54.39 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_resnet34.pth) |
|
| 41 |
+
| **DeFM ResNet-18** | 11.7 | 21.06 | 8.67 | 69.69 | 50.58 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_resnet18.pth) |
|
| 42 |
+
| **DeFM EfficientNet-B6** | 28.98 | 150.98 | 54.11 | 77.81 | 59.23 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_efficientnet_b6.pth) |
|
| 43 |
+
| **DeFM EfficientNet-B4** | 14.16 | 86.51 | 39.67 | 74.74 | 54.73 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_efficientnet_b4.pth) |
|
| 44 |
+
| **DeFM EfficientNet-B2** | 4.95 | 46.12 | 28.37 | 71.51 | 50.32 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_efficientnet_b2.pth) |
|
| 45 |
+
| **DeFM EfficientNet-B0** | 3.01 | 29.39 | 21.04 | 67.98 | 46.17 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_efficientnet_b0.pth) |
|
| 46 |
+
| **DeFM RegNetY-1.6GF** | 12.4 | 44.25 | 41.82 | 76.21 | 57.28 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_regnet_y_1_6gf.pth) |
|
| 47 |
+
| **DeFM RegNetY-800MF** | 6.3 | 25.21 | 24.16 | 74.91 | 57.03 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_regnet_y_800mf.pth) |
|
| 48 |
+
| **DeFM RegNetY-400MF** | 4.1 | 17.27 | 25.17 | 72.87 | 50.51 | [Download](https://huggingface.co/leggedrobotics/defm/resolve/main/defm_regnet_y_400mf.pth) |
|
| 49 |
+
---
|
| 50 |
+
|
| 51 |
+
## 📖 Citation
|
| 52 |
+
If you find DeFM useful for your research, please cite our paper:
|
| 53 |
+
|
| 54 |
+
```
|
| 55 |
+
@misc{patel2026defm,
|
| 56 |
+
title = {DeFM: Learning Foundation Representations from Depth for Robotics},
|
| 57 |
+
author = {Patel, Manthan and Frey, Jonas and Mittal, Mayank and Yang, Fan and Hansson, Alexander and Bar, Amir and Cadena, Cesar and Hutter, Marco},
|
| 58 |
+
year = {2026},
|
| 59 |
+
archivePrefix = {arXiv},
|
| 60 |
+
eprint = {XXXX.XXXXX},
|
| 61 |
+
primaryClass = {cs.RO}
|
| 62 |
+
}
|
| 63 |
+
```
|