poc

app.py

@@ -13,6 +13,11 @@ This is the demo for a Open Vocabulary Image Segmentation using
 [Segment Anything Model](https://github.com/facebookresearch/segment-anything) and
 [MetaCLIP](https://github.com/facebookresearch/MetaCLIP) combo.
 """
+EXAMPLES = [
+    ["https://media.roboflow.com/notebooks/examples/dog.jpeg", "dog"],
+    ["https://media.roboflow.com/notebooks/examples/dog.jpeg", "building"],
+    ["https://media.roboflow.com/notebooks/examples/dog-3.jpeg", "jacket"],
+]
 
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 SAM_GENERATOR = pipeline(
@@ -78,32 +83,44 @@ def filter_detections(
     return detections[filtering_mask]
 
 
-def inference(image_rgb_pil: Image.Image, prompt: str) -> Image.Image:
+def inference(image_rgb_pil: Image.Image, prompt: str) -> List[Image.Image]:
     width, height = image_rgb_pil.size
     area = width * height
 
     detections = run_sam(image_rgb_pil)
-    detections = detections[detections.area / area > 0.005]
+    detections = detections[detections.area / area > 0.01]
     detections = filter_detections(
         image_rgb_pil=image_rgb_pil,
         detections=detections,
         prompt=prompt)
 
-    return annotate(image_rgb_pil=image_rgb_pil, detections=detections)
+    return [
+        annotate(image_rgb_pil=image_rgb_pil, detections=detections),
+        annotate(image_rgb_pil=Image.new("RGB", (width, height), "black"), detections=detections)
+    ]
 
 
 with gr.Blocks() as demo:
     gr.Markdown(MARKDOWN)
     with gr.Row():
         with gr.Column():
-            input_image = gr.Image(image_mode='RGB', type='pil')
+            input_image = gr.Image(image_mode='RGB', type='pil', height=500)
             prompt_text = gr.Textbox(label="Prompt", value="dog")
-        result_image = gr.Image(image_mode='RGB', type='pil')
-    submit_button = gr.Button("Submit")
+            submit_button = gr.Button("Submit")
+        gallery = gr.Gallery(label="Result", object_fit="scale-down", preview=True)
+    with gr.Row():
+        gr.Examples(
+            examples=EXAMPLES,
+            fn=inference,
+            inputs=[input_image, prompt_text],
+            outputs=[gallery],
+            cache_examples=True,
+            run_on_click=True
+        )
 
     submit_button.click(
         inference,
         inputs=[input_image, prompt_text],
-        outputs=result_image)
+        outputs=gallery)
 
-demo.launch(debug=False)
+demo.launch(debug=False, show_error=True)