| --- |
| license: apache-2.0 |
| library_name: libreyolo |
| pipeline_tag: image-segmentation |
| tags: |
| - libreyolo |
| - picosam3 |
| - segmentation |
| - promptable-segmentation |
| - edge |
| - imx500 |
| --- |
| |
| # LibrePicoSAM3 |
|
|
| Edge tier of the LibreYOLO `LibreSAM` promptable-segmentation family: a |
| 1,371,418-parameter CNN that segments the object inside a box-defined region of |
| interest. It is small enough to run in-sensor on the Sony IMX500 and is the |
| smallest promptable segmenter LibreYOLO ships, sitting below MobileSAM. |
|
|
| ## Usage |
|
|
| ```python |
| from libreyolo import LibreSAM |
| |
| model = LibreSAM("picosam3") |
| result = model.predict("image.jpg", bboxes=[100, 100, 400, 500]) |
| result.show() |
| ``` |
|
|
| `bboxes=` is the only supported prompt. Point, text, mask, multimask and |
| segment-everything modes are not part of the upstream model contract and raise |
| a clear error; use `LibreSAM("sam2")` or `LibreSAM("sam3")` for those. |
|
|
| ## Measured quality |
|
|
| Evaluated by LibreYOLO on **COCO val2017**, using ground-truth boxes as ROI |
| prompts and ground-truth masks as reference: |
|
|
| | Protocol | mIoU | |
| |---|---| |
| | Crop space, 96x96 (upstream evaluation protocol) | **0.692** | |
| | Full image, end-to-end via `LibrePicoSAM3.predict()` | **0.697** | |
|
|
| Measured on 2000 randomly sampled non-crowd instances (seed 0, no area filter). |
| The PicoSAM3 paper reports 65.45 mIoU on COCO; our sampling and filtering may |
| differ from theirs, so we publish the number we measured rather than restating |
| theirs as reproduced. |
|
|
| ### Against MobileSAM |
|
|
| Same 150 COCO instances, same box prompts, single-threaded CPU: |
|
|
| | Model | Params | mIoU | CPU latency | |
| |---|---|---|---| |
| | LibrePicoSAM3 | 1.37M | 0.691 | **8.6 ms/prompt** | |
| | LibreMobileSAM | 10.13M | **0.800** | 407 ms/prompt | |
|
|
| MobileSAM is the better segmenter; prefer it when a GPU is available or quality |
| dominates. PicoSAM3 trades ~11 mIoU points for 7x fewer parameters and ~47x |
| lower CPU latency, which is what makes CPU-only and in-sensor deployment viable. |
|
|
| ## Provenance |
|
|
| - **Architecture**: ported from [pbonazzi/picosam3](https://github.com/pbonazzi/picosam3) |
| (Apache-2.0), commit `1b03949e43472953bb0021685c7fc3f5fdf48fde`. The LibreYOLO |
| implementation is native and produces bit-identical outputs to upstream |
| (max abs diff 0.0 on a seeded FP32 batch). |
| - **Weights**: converted from |
| [pietrobonazzi/picosam3](https://huggingface.co/pietrobonazzi/picosam3) |
| (Apache-2.0), revision `af49e4322b6b7cf448499fee5c073d4576f59444`, file |
| `PicoSAM3_SAM3_student_best.pt`. Re-wrapped into the LibreYOLO checkpoint |
| schema (v1) with provenance metadata; tensor values are unchanged. |
| - **Teacher chain**: upstream distilled this student from SAM 2.1 and SAM 3. |
| Teacher implementations and checkpoints are not included or redistributed |
| here. Meta's SAM License places ownership of distilled derivative works with |
| their creator and does not impose license inheritance, so these weights are |
| redistributed under the upstream author's Apache-2.0 terms. |
| - **Training data**: COCO 2017 and LVIS v1. |
|
|
| Note: the `PicoSAM3_student_epoch1.pt` and `PicoSAM3_epoch1.pt` files in the |
| upstream weights repo contain the older PicoSAM2 architecture (`output_head.*`) |
| and do not load as PicoSAM3. LibreYOLO ships the `best` artifact and rejects the |
| epoch-1 files explicitly. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{picosam3_2026, |
| title={PicoSAM3: Real-Time In-Sensor Region-of-Interest Segmentation}, |
| journal={IEEE Sensors Journal}, |
| year={2026} |
| } |
| ``` |
|
|
| ## License |
|
|
| Apache-2.0, inherited from the upstream code and weights. |
|
|
