Martin L (GitHub Actions) commited on
Commit
3d61e6f
·
1 Parent(s): 34ae677

Automated deployment from GitHub (source commit: 639602)

Browse files
Files changed (1) hide show
  1. docs/roadmap.md +27 -1
docs/roadmap.md CHANGED
@@ -1,7 +1,33 @@
1
- On this page, we outline the project's intended roadmap. This plan helps us strategically manage our time and resources.
2
 
3
  Below, we present our roadmap. It may evolve over time, so we will preserve previous versions to maintain transparency.
4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
  ## Roadmap as of *2025-05-03*
6
 
7
  ### Visual Overview
 
1
+ On this page, we outline the project's intended roadmap. The roadmap helps us strategically manage our time and resources.
2
 
3
  Below, we present our roadmap. It may evolve over time, so we will preserve previous versions to maintain transparency.
4
 
5
+ ## Roadmap as of *2025-11-08*
6
+
7
+ Our project currently faces two main challenges: deployment and prediction quality.
8
+
9
+ 1. **Deployment**: Since the plugin runs on the edge, the deployment is
10
+ really the installation on the user's device. This installation has
11
+ been a problem because we only tested it on Linux and the installation process
12
+ requires technical knowledge. These issues have recently been mitigated
13
+ through a new [online demo](https://huggingface.co/spaces/PellelNitram/xournalpp_htr)
14
+ that allows users to try Xournal++ HTR without installation. Following
15
+ the [successful PyInstaller proof of concept](PoCs/simplify_installation_with_single_file_executable.md),
16
+ this setup will likely become the preferred installation method in the future.
17
+
18
+ 2. **Prediction Quality**: We want to improve prediction quality. Improving prediction performance requires a way to measure real-life accuracy. To achieve this, two components are planned:
19
+
20
+ * A crowdsourced benchmark dataset, compiled from donated [demo](https://huggingface.co/spaces/PellelNitram/xournalpp_htr) input data.
21
+ * An end-to-end evaluation metric to assess both text detection and transcription quality.
22
+
23
+ Additionally, we currently rely on a third-party library called `htr_pipeline` for model delivery.
24
+ While effective, it complicates installation and model management. We plan to replace it with an
25
+ in-house implementation. The first component, the WordDetectorNet, has already been internalized
26
+ (see [explanation of model](https://lellep.xyz/blog/worddetectornet-visually-explained.html),
27
+ [code](https://github.com/PellelNitram/xournalpp_htr/tree/master/xournalpp_htr/training/WordDetectorNN) and
28
+ [demo](https://huggingface.co/spaces/PellelNitram/xournalpp_htr_WordDetectorNN)).
29
+ This transition away from `htr_pipeline` will not be backward compatible.
30
+
31
  ## Roadmap as of *2025-05-03*
32
 
33
  ### Visual Overview