| --- |
| title: SimamAnything2GS |
| emoji: ✨ |
| colorFrom: blue |
| colorTo: blue |
| sdk: gradio |
| app_file: app.py |
| --- |
| |
| # SimamAnything2GS |
|
|
| An ingest-first workbench for turning arbitrary images and videos into a |
| reproducible Gaussian-reconstruction job. The initial release deliberately |
| exports a CPU-only 2.5D billboard preview so the upload, hashing, and artifact |
| contract can be tested without model weights or GPU allocation. Images are used |
| directly; videos are sampled with FFmpeg (up to eight frames) and the first |
| sampled frame is used for the preview. |
|
|
| The preview is not a 3D reconstruction. Future adapters can add depth/pose |
| estimation, multi-view fusion, and Gaussian optimization |
| behind the same manifest. Each adapter must report its model, runtime, input |
| hashes, and held-out-view metrics. |
|
|
| `depth_bridge.py` now defines and tests the handoff contract for those adapters: |
| normalized depth plus optional confidence, intrinsics, and world-to-camera pose |
| become confidence-weighted, source-aware fused points. It performs no learned |
| inference, so the contract can be validated before a model backend is selected. |
|
|
| ## Verified smoke tests |
|
|
| The public Gradio API has been exercised with synthetic image and video inputs. |
| The video path sampled eight FFmpeg frames, exported 256 preview points, and |
| returned a manifest containing the video hash and sampling metadata. These are |
| ingest/export checks only, not reconstruction-quality results. |
|
|
