| --- |
| license: apache-2.0 |
| library_name: pytorch |
| pipeline_tag: image-segmentation |
| tags: |
| - fsot |
| - unet |
| - cell-detection |
| - zero-shot |
| - biomedical |
| --- |
| |
| # Wavegazer Net |
|
|
| Seed-derived visual U (FSOT) for segmentation and cell **centroid detect**. |
| **Zero trainable weights** on the spine. Pin `D1D38A`. |
|
|
| This is **not** FlowNet (optical flow, ICCV 2015). |
|
|
| ## Use |
|
|
| ```python |
| import torch |
| from wavegazer import WavegazerNet |
| |
| net = WavegazerNet(in_channels=1, n_classes=2, sparse=True).eval() |
| x = torch.rand(1, 1, 256, 256) |
| logits = net(x) # dense (N, K, H, W) |
| peaks = net.detect(x) # centroids, 7 µm Biohub protocol |
| ``` |
|
|
| Frozen codon kernels load with the module; optional `wavegazer_buffers.pt` |
| is the same buffers serialized. |
|
|
| ## Benchmarks (honest) |
|
|
| - Dense square unit test: Dice > 0.7 |
| - Synthetic disks vs trained U-Net: U-Net wins (fitted); Wavegazer is closed-form |
| - Biohub detect @ 7 µm (16 volumes, YX): see `artifacts/biohub_peaks_7um.json` |
|
|
| Not a CellMot 0.848 claim. Detect gate first, linking later. |
|
|
| Code: GitHub `dappalumbo91/Wavegazer-Net` |
|
|
