Update RGB image format from JPG to PNG and enhance Gradio UI with troubleshooting tips

app.py

@@ -24,12 +24,10 @@ from masks import generate_naive_mask
 
 DEFAULT_DATA_DIR = Path("/app/tests/reference/t_shape")
 DEFAULT_MESH = DEFAULT_DATA_DIR / "t_shape.obj"
-DEFAULT_RGB = DEFAULT_DATA_DIR / "rgb_001.jpg"
+DEFAULT_RGB = DEFAULT_DATA_DIR / "rgb_001.png"
 DEFAULT_DEPTH = DEFAULT_DATA_DIR / "depth_001.png"
 DEFAULT_REF_IMAGES = [
-    DEFAULT_DATA_DIR / "rgb_001.jpg",
-    DEFAULT_DATA_DIR / "rgb_002.jpg",
-    DEFAULT_DATA_DIR / "rgb_003.jpg",
+    DEFAULT_DATA_DIR / "rgb_001.png"
 ]
 
 _slimsam_model = None
@@ -523,6 +521,7 @@ def gradio_estimate(
 # Gradio UI
 with gr.Blocks(title="FoundationPose Inference", theme=gr.themes.Soft()) as demo:
     gr.Markdown("# 🎯 FoundationPose 6D Object Pose Estimation")
+    gr.Markdown("Project page: https://nvlabs.github.io/FoundationPose/")
 
     mode_indicator = gr.Markdown(
         "**Mode:** 🟢 Real FoundationPose with GPU",
@@ -599,6 +598,12 @@ with gr.Blocks(title="FoundationPose Inference", theme=gr.themes.Soft()) as demo
             Upload a query image containing the initialized object.
             The model will estimate the 6D pose (position + orientation).
             """)
+            gr.Markdown("""
+            **Troubleshooting**
+            - **Camera intrinsics**: make sure `fx`, `fy`, `cx`, `cy` match the query RGB resolution. If you resize the RGB image, scale intrinsics by the same factor.
+            - **Depth values**: verify units. 16-bit PNG is usually millimeters (converted to meters here). If values look clipped or too small/large, confirm the sensor's depth scale.
+            - **Image scales**: RGB and depth must be the same size. If they differ, depth is resized to match RGB (nearest-neighbor). Prefer exporting aligned pairs from the same stream.
+            """)
 
             with gr.Row():
                 with gr.Column():