| --- |
| library_name: pytorch |
| --- |
| |
|  |
|
|
| PSCVStereo is a lightweight end-to-end neural stereo matching model inspired by RAFT-Stereo, IGEV-Stereo and CREStereo, optimized for deployment on Ambarella CVflow hardware (CV7, CV72, and N1-655). It estimates pixel-wise disparity from left and right camera images for depth perception on robotics and autonomous mobile robots (AMRs). |
| - **Technical**: Uses a lightweight correlation cost volume and iterative refinement, with CVflow-aware optimizations and sparsification. |
| - **Applications**: Depth estimation, obstacle detection, and navigation for edge robotics and AMRs. |
| - **Output**: Disparity map and confidence mask for confidence-aware depth perception. |
|
|
| **Related works**: |
| [RAFT-Stereo: Multilevel Recurrent Field Transforms for Stereo Matching](https://arxiv.org/abs/2109.07547) |
| [Iterative Geometry Encoding Volume for Stereo Matching (IGEV-Stereo)](https://arxiv.org/abs/2303.06615) |
| [Practical Stereo Matching via Cascaded Recurrent Network (CREStereo)](https://arxiv.org/abs/2203.11483) |
| **Reference implementation**: |
| [RAFT-Stereo GitHub repository](https://github.com/princeton-vl/RAFT-Stereo) |
| [IGEV-Stereo GitHub repository](https://github.com/gangweiX/IGEV) |
| [CREStereo GitHub repository](https://github.com/megvii-research/CREStereo) |
| |
| # PSCVStereo-base |
|
|
| This model is the higher-capacity model. It uses a full-width feature backbone and a richer correlation volume, refined with more coarse-to-fine updates over a wider disparity range. It is trained on 512×896, so it is the choice when accuracy on larger images matters most. |
|
|
| Model Configuration: |
| - Implementation: [PSCVStereo](https://github.com/Ambarella-Inc/cooper-pscvstereo) |
| - Resolution: 3x512x896 |
| - Support Cooper version: |
| - Cooper SDK: [3.0.0] |
| - Cooper Foundry: [2.3] |
|
|
| | Model | Device | compression | Model Link | |
| | :-----: | :-----: | :-----: | ------- | |
| | PSCVStereo-base | N1-655 | Amba_optimized | [Model_Link](https://huggingface.co/Ambarella/PSCVStereo/blob/main/n1-655_pscvstereo_base_amba_optimized.bin) | |
| | PSCVStereo-base | CV7 | Amba_optimized | [Model_Link](https://huggingface.co/Ambarella/PSCVStereo/blob/main/cv7_pscvstereo_base_amba_optimized.bin) | |
| | PSCVStereo-base | CV72 | Amba_optimized | [Model_Link](https://huggingface.co/Ambarella/PSCVStereo/blob/main/cv72_pscvstereo_base_amba_optimized.bin) | |
| |
| # PSCVStereo-tiny |
| |
| This model is a more compact variant. It uses a narrower backbone and a slimmer correlation volume, with fewer iterative updates. It is trained on 416×576, so it is the choice when lower compute cost and faster inference are the priority. |
| |
| Model Configuration: |
| - Implementation: [PSCVStereo](https://github.com/Ambarella-Inc/cooper-pscvstereo) |
| - Resolution: 3x416x576 |
| - Support Cooper version: |
| - Cooper SDK: [3.0.0] |
| - Cooper Foundry: [2.3] |
| |
| | Model | Device | compression | Model Link | |
| | :-----: | :-----: | :-----: | ------- | |
| | PSCVStereo-tiny | N1-655 | Amba_optimized | [Model_Link](https://huggingface.co/Ambarella/PSCVStereo/blob/main/n1-655_pscvstereo_tiny_amba_optimized.bin) | |
| | PSCVStereo-tiny | CV7 | Amba_optimized | [Model_Link](https://huggingface.co/Ambarella/PSCVStereo/blob/main/cv7_pscvstereo_tiny_amba_optimized.bin) | |
| | PSCVStereo-tiny | CV72 | Amba_optimized | [Model_Link](https://huggingface.co/Ambarella/PSCVStereo/blob/main/cv72_pscvstereo_tiny_amba_optimized.bin) | |
|
|
