Upload README.md with huggingface_hub

README.md

@@ -32,7 +32,8 @@ Arabic are not covered.
 | `ppocr_rec.yml` | Recognizer character dictionary |
 | `ppocr_det_dynamic.onnx` | fp32 detection, CPU reference for accuracy checks |
 | `ppocr_rec_dynamic.onnx` | fp32 recognition, CPU reference for accuracy checks |
-| `samples/sample.png` | Sample document page |
+| `samples/sample.jpg` | Sample café menu card, 10 text lines |
+| `samples/sample.png` | Sample document page, 99 text lines |
 
 ### Why four recognition models
 
@@ -49,15 +50,15 @@ python setup_demos.py ppocr
 
 cd ppocr
 python src/infer.py \
-  --image  ../models/Synaptics/paddle-paddle-tiny/samples/sample.png \
+  --image  ../models/Synaptics/paddle-paddle-tiny/samples/sample.jpg \
   --models ../models/Synaptics/paddle-paddle-tiny \
-  --tda cpu \
   --save-image
 ```
 
-> On current firmware the detection model can fail with
-> `INTERNAL; failed to writeXram()` under the default `dmabuf` allocator. Pass
-> `--tda cpu` if you hit this.
+> The demo defaults to `--tda cpu`, the Torq runtime's own default allocator.
+> Do not switch it to `dmabuf`: on current firmware the detection model fails
+> there with `INTERNAL; failed to writeXram()`. This selects the allocator
+> backing device buffers only — both stages run on the NPU either way.
 
 Either stage can be switched to ONNX Runtime with `--det-backend ort` /
 `--rec-backend ort` (plus the matching `--det-onnx` / `--rec-onnx`) to compare
@@ -65,12 +66,34 @@ NPU output against a CPU reference.
 
 ## Measured on SL2619
 
-Sample page, 1224×1584, 99 text lines detected:
+| Sample | Lines | Detection | Recognition |
+|---|---|---|---|
+| `sample.jpg` (menu card, 912×1200) | 10 | ~0.52 s | ~1.2 s |
+| `sample.png` (paper page, 1224×1584) | 99 | ~0.70 s | ~22.2 s |
 
-| Stage | Time |
-|---|---|
-| Detection (800×608) | ~0.70 s |
-| Recognition (99 lines, bucketed) | ~22.2 s |
+Recognition scales with line count because each line is a separate invocation —
+the bucket models are compiled with a static batch of 1.
+
+### About `sample.jpg`
+
+A 912×1200 café menu card, rendered synthetically in DejaVu Serif rather than
+photographed, so it carries no third-party image licensing. Two properties of it
+are worth copying if you make your own sample:
+
+- **Aspect ratio 0.76** (912/1200), matching the detector's static 608×800
+  input. Preprocessing resizes straight to that shape without preserving
+  aspect, so an off-ratio image is fed to the model stretched.
+- **Prices sit two spaces after the item name**, not right-aligned with a dot
+  leader. DBNet does not bridge a wide intra-line gap — with right-aligned
+  prices each row detects as two boxes (name and price separately), nearly
+  doubling the line count and the recognition time.
 
-Recognition dominates because each line is a separate invocation — the bucket
-models are compiled with a static batch of 1.
+Recognized output, all ten lines at confidence ≥ 0.966:
+
+```
+  1   [0.991] BLUE DOOR CAFE          6   [0.995] Smoked Salmon Bagel  9.75
+  2   [0.996] all day breakfast       7   [1.000] DRINKS
+  3   [0.999] BREAKFAST               8   [0.975] Espresso  2.75
+  4   [0.996] Avocado Toast  6.50     9   [0.966] Fresh Orange Juice  4.00
+  5   [0.996] Buttermilk Pancakes  7.00  10  [0.993] open 7am - 3pm daily
+```