| --- |
| license: mit |
| library_name: pytorch |
| pipeline_tag: depth-estimation |
| tags: |
| - monocular-depth-estimation |
| - depth-estimation |
| - computer-vision |
| - pytorch |
| - mobilenetv2 |
| - knowledge-distillation |
| --- |
| |
| # ANCHURUS |
|
|
| ### Lightweight Monocular Depth Estimation |
|
|
| **ANCHURUS** is an experimental deep learning model for monocular depth estimation, developed by Davitestro. |
|
|
| The project explores computationally efficient depth prediction using a student-teacher learning approach. |
|
|
| Its primary goal is to investigate lightweight neural network architectures capable of generating depth maps from single RGB images while keeping computational requirements relatively low. |
|
|
| ## Model Details |
|
|
| | Property | Value | |
| |---|---| |
| | Model | ANCHURUS v5 | |
| | Framework | PyTorch | |
| | Task | Monocular Depth Estimation | |
| | Encoder | MobileNetV2 | |
| | Architecture | Encoder-Decoder | |
| | Input | RGB, 256 × 256 | |
| | Output | Single-channel depth map | |
| | Output range | 0–1 (normalized) | |
| | Checkpoint | `model/ANCHURUS-v5.pth` | |
| | License | MIT | |
|
|
| ## Architecture |
|
|
| ANCHURUS uses a MobileNetV2 feature extraction backbone followed by a progressive decoder consisting of five upsampling blocks. |
|
|
| The network produces a normalized depth representation from a single RGB image. |
|
|
| ```text |
| RGB Image |
| | |
| v |
| MobileNetV2 Encoder |
| | |
| v |
| Feature Maps |
| | |
| v |
| Progressive Decoder |
| (5 Upsampling Blocks) |
| | |
| v |
| Sigmoid Activation |
| | |
| v |
| Normalized Depth Map |
| ``` |
|
|
| ## Training Approach |
|
|
| ANCHURUS was developed using a student-teacher learning framework. |
|
|
| MiDaS serves as the teacher model, providing depth-related supervision for the lightweight student architecture. |
|
|
| The training implementation also includes an adaptive learning-rate strategy. |
|
|
| ## Model Weights |
|
|
| The recommended checkpoint is: |
|
|
| **[ANCHURUS v5 — Download](https://huggingface.co/Testro/ANCHURUS/resolve/main/model/ANCHURUS-v5.pth)** |
|
|
| The checkpoint contains the model state dictionary. |
|
|
| ## Intended Applications |
|
|
| Potential research and development applications include: |
|
|
| - Robotics and computer vision. |
| - Embedded perception systems. |
| - Resource-constrained inference. |
| - Monocular depth estimation experiments. |
| - Real-time visual perception research. |
|
|
| These applications are not claims of validated real-world performance. |
|
|
| ## Limitations |
|
|
| ANCHURUS is an experimental model. |
|
|
| Its output represents normalized depth values and should not be interpreted as calibrated metric distance without additional calibration. |
|
|
| Prediction quality and generalization depend on training data and environmental conditions. |
|
|
| Independent benchmarking is encouraged. |
|
|
| ## Source Code |
|
|
| The complete implementation, training scripts and development history are available on GitHub: |
|
|
| **[Davitestro/ANCHURUS](https://github.com/Davitestro/ANCHURUS)** |
|
|
| ## Author |
|
|
| **Davitestro** |
|
|
| GitHub: https://github.com/Davitestro |
|
|
| Hugging Face: https://huggingface.co/Testro |
|
|
| ## License |
|
|
| MIT License. |
|
|
| Copyright (c) 2026 Davit |
|
|
| See the [LICENSE](LICENSE) file for the complete license text. |
|
|
| Feedback, independent evaluation and research collaboration are welcome. |
|
|
