first commit

README.md

@@ -1,3 +1,135 @@
 ---
 license: apache-2.0
+task_categories:
+  - image-to-image
+language:
+  - en
+  - zh
+tags:
+  - VAE
+  - text-rich document
+  - text-reconstruction
+  - benchmark
+size_categories:
+  - 1K<n<10K
+pretty_name: OmniDoc-TokenBench
 ---
+
+# OmniDoc-TokenBench
+
+Technical report and dataset are coming soon.
+
+---
+
+## Introduction
+
+**OmniDoc-TokenBench** is a curated benchmark of ~3K text-rich document images for evaluating VAE reconstruction on textual content, alongside an evaluation toolkit supporting PSNR, SSIM, LPIPS, FID, and OCR-based NED metrics. It spans nine categories (*book*, *slides*, *color textbook*, *exam paper*, *academic paper*, *magazine*, *financial report*, *newspaper*, *note*) in both English and Chinese.
+
+<p align="center">
+  <img src="assets/bench.png" alt="OmniDoc-TokenBench" style="max-width: 1328px; width: 100%; height: auto;" />
+<p>
+
+Derived from OmniDocBench, each sample is cropped from a text block and resized to 256x256 with reference character sizes of 16px (Chinese) and 10px (English). We filter for sufficient character density ([200, 600] for Chinese, [300, 600] for English), deduplicate via n-gram overlap, and manually inspect for quality.
+
+**Evaluation.** Beyond standard traditional metrics (PSNR, SSIM, LPIPS, FID), we use **NED** (Normalized Edit Distance) as the primary text-fidelity metric. NED compares the OCR outputs of the original and reconstructed images using Levenshtein distance:
+
+$$
+\mathrm{NED} = \frac{1}{N}\sum_{i=1}^{N}\left(1 - \frac{d_{\mathrm{edit}}(s_{\mathrm{gt}}^{(i)}, s_{\mathrm{recon}}^{(i)})}{\max(|s_{\mathrm{gt}}^{(i)}|, |s_{\mathrm{recon}}^{(i)}|)}\right)
+$$
+
+---
+
+## Results
+
+We conduct a comprehensive evaluation on OmniDoc-TokenBench (~3K text-rich images, 256×256). Models are grouped by spatial compression factor and sorted by NED within each group.
+
+<p align="center">
+  <img src="assets/results.png" alt="Eval-Results" style="max-width: 1024px; width: 100%; height: auto;" />
+<p>
+
+---
+
+## Evaluation
+
+### Installation
+
+```bash
+pip install torch torchvision piq lpips pytorch-fid pillow numpy tqdm
+pip install paddleocr python-Levenshtein  # required for NED
+```
+
+### Usage
+
+Place your ground-truth images in `gt_dir/` and reconstructed images in `recon_dir/` (filenames must match one-to-one).
+
+```bash
+# Compute NED only (default)
+python eval_metrics.py --gt_dir ./gt_dir --recon_dir ./recon_dir
+
+# Compute traditional metrics (PSNR / SSIM / LPIPS / FID)
+python eval_metrics.py --gt_dir ./gt_dir --recon_dir ./recon_dir --mode pixel
+
+# Compute all metrics
+python eval_metrics.py --gt_dir ./gt_dir --recon_dir ./recon_dir --mode all
+
+# Specify output directory and device
+python eval_metrics.py --gt_dir ./gt_dir --recon_dir ./recon_dir --save_path ./results --device cuda
+```
+
+### Output
+
+The script writes results to the `--save_path` directory (default: `./eval_results`):
+
+- `results.json` --- Aggregated metrics (example):
+  ```json
+  {
+    "num_samples": 100,
+    "PSNR": 30.45,
+    "SSIM": 0.9706,
+    "LPIPS": 0.0523,
+    "FID": 1.98,
+    "NED": 0.9617,
+    "NED_samples": 98
+  }
+  ```
+
+- `ned_details.json` --- Per-image OCR results and NED scores (generated when `--mode` is `ned` or `all`):
+  ```json
+  {
+    "avg_ned": 0.9617,
+    "total_samples": 100,
+    "valid_samples": 98,
+    "details": [
+      {
+        "file": "0001.png",
+        "gt_ocr": "The quick brown fox...",
+        "recon_ocr": "The quick brown fox...",
+        "ned": 0.9764
+      }
+    ]
+  }
+  ```
+
+### Notes
+
+- `FID` and `LPIPS` require downloading model checkpoints on the first run (InceptionV3 ~90MB for FID, VGG16 ~530MB for LPIPS). Ensure network access or pre-download the weight files.
+- PaddleOCR defaults to CPU inference. For large-scale evaluation, consider switching to GPU by setting `device="gpu"` in `compute_ned()`.
+- The progress bar for PSNR/SSIM/LPIPS displays running means in real time.
+
+---
+
+## Citation
+
+If you use OmniDoc-TokenBench or this evaluation toolkit in your research, please cite:
+
+```bibtex
+% TODO
+```
+
+---
+
+## License
+
+- OmniDoc-TokenBench is a derivative dataset based on [OmniDocBench](https://github.com/opendatalab/OmniDocBench), developed by the Qwen Team at Alibaba Group.
+
+- This project is licensed under the [Apache License 2.0](LICENSE).