Card: link to the new stroke checks in the GitHub README
Browse files
README.md
CHANGED
|
@@ -1,28 +1,29 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: mit
|
| 3 |
-
tags:
|
| 4 |
-
- vesuvius-challenge
|
| 5 |
-
- ink-detection
|
| 6 |
-
- herculaneum
|
| 7 |
-
- ct
|
| 8 |
-
---
|
| 9 |
-
|
| 10 |
-
# Reader V3
|
| 11 |
-
|
| 12 |
-
**Clearer letters on scrolls it has never seen: strokes that Reader v2 breaks apart and other 9 µm models blur into
|
| 13 |
-
texture or speckle come out whole and separate, with over 10x less false ink than Hecate, and the fastest of them all.**
|
| 14 |
-
|
| 15 |
-
Reader V3 is the upgrade to [Reader v2](https://huggingface.co/domenicor046/reader-v2): a drop-in checkpoint for the
|
| 16 |
-
Vesuvius Challenge team's own `ink_9um` inference command (`koine_machines.inference.infer`), for the native 9.36 µm and
|
| 17 |
-
8.64 µm scans. Every result is on data the model never trained on.
|
| 18 |
-
|
| 19 |
-

|
| 20 |
-
|
| 21 |
-
```bash
|
| 22 |
-
python -m koine_machines.inference.infer <surface_volume.zarr> reader-v3.pth out.tif --overlap 0.5 --blend-mode hann --direction both
|
| 23 |
-
```
|
| 24 |
-
|
| 25 |
-
Full results, whole never-seen segments, U08 on PHerc0343 (letters found by my own pipeline, shown for the first time
|
| 26 |
-
anywhere), the head-to-head
|
| 27 |
-
|
| 28 |
-
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- vesuvius-challenge
|
| 5 |
+
- ink-detection
|
| 6 |
+
- herculaneum
|
| 7 |
+
- ct
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# Reader V3
|
| 11 |
+
|
| 12 |
+
**Clearer letters on scrolls it has never seen: strokes that Reader v2 breaks apart and other 9 µm models blur into
|
| 13 |
+
texture or speckle come out whole and separate, with over 10x less false ink than Hecate, and the fastest of them all.**
|
| 14 |
+
|
| 15 |
+
Reader V3 is the upgrade to [Reader v2](https://huggingface.co/domenicor046/reader-v2): a drop-in checkpoint for the
|
| 16 |
+
Vesuvius Challenge team's own `ink_9um` inference command (`koine_machines.inference.infer`), for the native 9.36 µm and
|
| 17 |
+
8.64 µm scans. Every result is on data the model never trained on.
|
| 18 |
+
|
| 19 |
+

|
| 20 |
+
|
| 21 |
+
```bash
|
| 22 |
+
python -m koine_machines.inference.infer <surface_volume.zarr> reader-v3.pth out.tif --overlap 0.5 --blend-mode hann --direction both
|
| 23 |
+
```
|
| 24 |
+
|
| 25 |
+
Full results, whole never-seen segments, U08 on PHerc0343 (letters found by my own pipeline, shown for the first time
|
| 26 |
+
anywhere), the head-to-head, three checks that the larger context does not invent strokes or close letters (a c into an
|
| 27 |
+
o), and a script that checks every number: https://github.com/DomRusso2/reader-v3
|
| 28 |
+
|
| 29 |
+
MIT licence. Author: Domenico Russo.
|