Prepare job build context

Dockerfile

@@ -8,5 +8,6 @@ ENV USE_REAL_MODEL=true
 # Copy application files
 WORKDIR /app
 COPY app.py client.py estimator.py masks.py .
+COPY tests/reference/t_shape /app/tests/reference/t_shape
 
 CMD ["python3", "app.py"]

app.py

@@ -22,6 +22,16 @@ import torch
 
 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_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",
+]
+
 _slimsam_model = None
 _slimsam_processor = None
 _slimsam_device = None
@@ -510,13 +520,15 @@ with gr.Blocks(title="FoundationPose Inference", theme=gr.themes.Soft()) as demo
                             cad_mesh_file = gr.File(
                                 label="3D Mesh File (.obj, .stl, .ply)",
                                 file_count="single",
-                                file_types=[".obj", ".stl", ".ply", ".mesh"]
+                                file_types=[".obj", ".stl", ".ply", ".mesh"],
+                                value=str(DEFAULT_MESH) if DEFAULT_MESH.exists() else None
                             )
 
                             cad_ref_files = gr.File(
                                 label="Reference Images (Optional)",
                                 file_count="multiple",
-                                file_types=["image"]
+                                file_types=["image"],
+                                value=[str(p) for p in DEFAULT_REF_IMAGES if p.exists()]
                             )
 
                             gr.Markdown("### Camera Intrinsics")
@@ -561,12 +573,14 @@ with gr.Blocks(title="FoundationPose Inference", theme=gr.themes.Soft()) as demo
 
                     est_query_image = gr.Image(
                         label="Query Image (RGB)",
-                        type="numpy"
+                        type="numpy",
+                        value=str(DEFAULT_RGB) if DEFAULT_RGB.exists() else None
                     )
 
                     est_depth_image = gr.Image(
                         label="Depth Image (Optional, 16-bit PNG)",
-                        type="numpy"
+                        type="numpy",
+                        value=str(DEFAULT_DEPTH) if DEFAULT_DEPTH.exists() else None
                     )
 
                     est_mask_method = gr.Radio(

estimator.py

@@ -245,7 +245,9 @@ class FoundationPoseEstimator:
                     iteration=2  # Fewer iterations for tracking
                 )
 
-            # Store pose for next frame
+            # Store pose for next frame (move to CPU if it's a tensor)
+            if torch.is_tensor(pose):
+                pose = pose.detach().cpu().numpy()
             obj_data["pose_last"] = pose
 
             if pose is None: