Upload app.py

app.py

@@ -499,6 +499,70 @@ def run_segformer(img_rgb, morph_cleanup=False):
     except Exception as e:
         return img_rgb, None, f"SegFormer Error: {e}"
 
+# ═══════════════════════════════════════════════════════════════════════════════
+# BiRefNet Function
+# ═══════════════════════════════════════════════════════════════════════════════
+def run_birefnet(img_rgb):
+    try:
+        from transformers import AutoModelForImageSegmentation
+        from torchvision import transforms
+        import torch.nn.functional as F
+        
+        t0 = time.time()
+        
+        base_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+        
+        # Paths to try (works for local PC and Hugging Face Cloud deployment)
+        paths_to_try = [
+            os.path.join(base_dir, "BiRefNet_Model", "best_model-20260624T051601Z-3-001", "best_model"), # Local PC
+            "birefnet_model",                                                                          # Hugging Face Root / Root dir
+            os.path.join(os.path.dirname(os.path.abspath(__file__)), "birefnet_model"),                # Next to app.py
+            "best_birefnet_model"                                                                      # Extra fallback
+        ]
+        
+        model_path = None
+        for p in paths_to_try:
+            if os.path.exists(p) and os.path.exists(os.path.join(p, "config.json")):
+                model_path = p
+                break
+                
+        # Final fallback: Download directly from Hugging Face Model Repo!
+        if model_path is None:
+            model_path = "Ayesha-Majeed/birefnet_car_window" 
+            
+        model = AutoModelForImageSegmentation.from_pretrained(model_path, trust_remote_code=True).to(DEVICE)
+        model.eval()
+        
+        image_transform = transforms.Compose([
+            transforms.Resize((1024, 1024)),
+            transforms.ToTensor(),
+            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ])
+        
+        from PIL import Image
+        pil_img = Image.fromarray(img_rgb)
+        input_tensor = image_transform(pil_img).unsqueeze(0).to(DEVICE)
+        
+        with torch.no_grad():
+            if DEVICE == "cuda":
+                with torch.amp.autocast("cuda"):
+                    preds = model(input_tensor)
+                    final_pred = preds[-1] if isinstance(preds, (list, tuple)) else preds
+            else:
+                preds = model(input_tensor)
+                final_pred = preds[-1] if isinstance(preds, (list, tuple)) else preds
+                
+        h, w = img_rgb.shape[:2]
+        final_pred = F.interpolate(final_pred, size=(h, w), mode="bilinear", align_corners=False)
+        pred_mask = (torch.sigmoid(final_pred) > 0.5).squeeze().cpu().numpy().astype(np.uint8)
+        
+        elapsed = time.time() - t0
+        out = apply_mask_overlay(img_rgb, pred_mask > 0, color=(255, 0, 0)) # Red
+        bw_mask = (pred_mask * 255).astype(np.uint8)
+        return out, bw_mask, f"Found: 1 (Semantic) | Inference: {elapsed:.2f}s"
+    except Exception as e:
+        return img_rgb, None, f"BiRefNet Error: {e}"
+
 # ═══════════════════════════════════════════════════════════════════════════════
 # Gradio Process Function
 # ═══════════════════════════════════════════════════════════════════════════════
@@ -582,22 +646,23 @@ with gr.Blocks(theme=theme, title="Car Window Segmentation") as demo:
 
 **Custom Trained Models:**
 
-1. YOLOv8x-seg
+1. SegFormer
 2. YOLO11x-seg
-3. Mask R-CNN
-4. SegFormer
-5. SegFormer + Morphological
-6. SAM + YOLO (Strategy 1: Bbox + 5 Points)
-7. SAM + YOLO (Strategy 2: Mask + 5 Points)
-8. SAM + YOLO (Strategy 3: Direct Mask Prompting)
+3. BiRefNet
+4. YOLOv8x-seg
+5. Mask R-CNN
+6. SegFormer + Morphological
+7. SAM + YOLO (Strategy 1: Bbox + 5 Points)
+8. SAM + YOLO (Strategy 2: Mask + 5 Points)
+9. SAM + YOLO (Strategy 3: Direct Mask Prompting)
 
 **Pretrained Zero-Shot Models:**
 
-9. Grounding DINO
-10. Grounded SAM
-11. Intelliarts Car Parts
+10. Grounding DINO
+11. Grounded SAM
+12. Intelliarts Car Parts
 
-**Our Findings:** SegFormer and YOLO11x deliver the best performance with significantly sharper edge precision.
+**Our Findings:** BiRefNet, SegFormer, and YOLO11x deliver the best performance with significantly sharper edge precision.
 """)
         
         with gr.Row():
@@ -614,12 +679,12 @@ with gr.Blocks(theme=theme, title="Car Window Segmentation") as demo:
         gr.Markdown("---")
         gr.Markdown("## 🚀 Custom Trained Models")
         
-        gr.Markdown("### 1️⃣ YOLOv8x-seg")
+        gr.Markdown("### 1️⃣ SegFormer (Transformer)")
         with gr.Row():
-            seq_yolo_img = gr.Image(label="YOLO Overlay", interactive=False)
-            seq_yolo_bw = gr.Image(label="YOLO Binary Mask", interactive=False, image_mode="L")
-        seq_yolo_stats = gr.Textbox(label="YOLO Stats", interactive=False)
-            
+            seq_segf_img = gr.Image(label="SegFormer Overlay", interactive=False)
+            seq_segf_bw = gr.Image(label="SegFormer Binary Mask", interactive=False, image_mode="L")
+        seq_segf_stats = gr.Textbox(label="SegFormer Stats", interactive=False)
+
         gr.Markdown("---")
         gr.Markdown("### 2️⃣ YOLO11x-seg")
         with gr.Row():
@@ -628,42 +693,49 @@ with gr.Blocks(theme=theme, title="Car Window Segmentation") as demo:
         seq_yolo11_stats = gr.Textbox(label="YOLO11x Stats", interactive=False)
 
         gr.Markdown("---")
-        gr.Markdown("### 3️⃣ Mask R-CNN (ResNet50-FPN)")
+        gr.Markdown("### 3️⃣ BiRefNet (Boundary-Aware Model)")
         with gr.Row():
-            seq_mrcnn_img = gr.Image(label="Mask R-CNN Overlay", interactive=False)
-            seq_mrcnn_bw = gr.Image(label="Mask R-CNN Binary Mask", interactive=False, image_mode="L")
-        seq_mrcnn_stats = gr.Textbox(label="Mask R-CNN Stats", interactive=False)
+            seq_biref_img = gr.Image(label="BiRefNet Overlay", interactive=False)
+            seq_biref_bw = gr.Image(label="BiRefNet Binary Mask", interactive=False, image_mode="L")
+        seq_biref_stats = gr.Textbox(label="BiRefNet Stats", interactive=False)
+            
+        gr.Markdown("---")
+        gr.Markdown("### 4️⃣ YOLOv8x-seg")
+        with gr.Row():
+            seq_yolo_img = gr.Image(label="YOLO Overlay", interactive=False)
+            seq_yolo_bw = gr.Image(label="YOLO Binary Mask", interactive=False, image_mode="L")
+        seq_yolo_stats = gr.Textbox(label="YOLO Stats", interactive=False)
 
         gr.Markdown("---")
-        gr.Markdown("### 4️⃣ SegFormer (Transformer - Best Model)")
+        gr.Markdown("### 5️⃣ Mask R-CNN (ResNet50-FPN)")
         with gr.Row():
-            seq_segf_img = gr.Image(label="SegFormer Overlay", interactive=False)
-            seq_segf_bw = gr.Image(label="SegFormer Binary Mask", interactive=False, image_mode="L")
-        seq_segf_stats = gr.Textbox(label="SegFormer Stats", interactive=False)
+            seq_mrcnn_img = gr.Image(label="Mask R-CNN Overlay", interactive=False)
+            seq_mrcnn_bw = gr.Image(label="Mask R-CNN Binary Mask", interactive=False, image_mode="L")
+        seq_mrcnn_stats = gr.Textbox(label="Mask R-CNN Stats", interactive=False)
 
         gr.Markdown("---")
-        gr.Markdown("### 5️⃣ SegFormer + Morphological Cleanup (Holes Filled + Sharp Borders)")
+        gr.Markdown("### 6️⃣ SegFormer + Morphological Cleanup (Holes Filled + Sharp Borders)")
         with gr.Row():
             seq_segf_morph_img = gr.Image(label="SegFormer + Morph Overlay", interactive=False)
             seq_segf_morph_bw = gr.Image(label="SegFormer + Morph Binary Mask", interactive=False, image_mode="L")
         seq_segf_morph_stats = gr.Textbox(label="SegFormer + Morph Stats", interactive=False)
 
         gr.Markdown("---")
-        gr.Markdown("### 6️⃣ SAM + YOLO (Strategy 1: Bbox + 5 Points)")
+        gr.Markdown("### 7️⃣ SAM + YOLO (Strategy 1: Bbox + 5 Points)")
         with gr.Row():
             seq_sam1_img = gr.Image(label="SAM+YOLO Strat 1 Overlay", interactive=False)
             seq_sam1_bw = gr.Image(label="SAM+YOLO Strat 1 Binary Mask", interactive=False, image_mode="L")
         seq_sam1_stats = gr.Textbox(label="SAM+YOLO Strat 1 Stats", interactive=False)
 
         gr.Markdown("---")
-        gr.Markdown("### 7️⃣ SAM + YOLO (Strategy 2: Mask + 5 Points)")
+        gr.Markdown("### 8️⃣ SAM + YOLO (Strategy 2: Mask + 5 Points)")
         with gr.Row():
             seq_sam2_img = gr.Image(label="SAM+YOLO Strat 2 Overlay", interactive=False)
             seq_sam2_bw = gr.Image(label="SAM+YOLO Strat 2 Binary Mask", interactive=False, image_mode="L")
         seq_sam2_stats = gr.Textbox(label="SAM+YOLO Strat 2 Stats", interactive=False)
 
         gr.Markdown("---")
-        gr.Markdown("### 8️⃣ SAM + YOLO (Strategy 3: Direct Mask Prompting)")
+        gr.Markdown("### 9️⃣ SAM + YOLO (Strategy 3: Direct Mask Prompting)")
         with gr.Row():
             seq_sam3_img = gr.Image(label="SAM+YOLO Strat 3 Overlay", interactive=False)
             seq_sam3_bw = gr.Image(label="SAM+YOLO Strat 3 Binary Mask", interactive=False, image_mode="L")
@@ -672,21 +744,21 @@ with gr.Blocks(theme=theme, title="Car Window Segmentation") as demo:
         gr.Markdown("---")
         gr.Markdown("## 🌍 Pretrained Zero-Shot Models")
 
-        gr.Markdown("### 9️⃣ Grounding DINO (Zero-Shot Detection)")
+        gr.Markdown("### 🔟 Grounding DINO (Zero-Shot Detection)")
         with gr.Row():
             seq_dino_img = gr.Image(label="Grounding DINO Overlay", interactive=False)
             seq_dino_bw = gr.Image(label="Grounding DINO Binary Mask", interactive=False, image_mode="L")
         seq_dino_stats = gr.Textbox(label="Grounding DINO Stats", interactive=False)
 
         gr.Markdown("---")
-        gr.Markdown("### 🔟 Grounded SAM (Zero-Shot Segmentation)")
+        gr.Markdown("### 1️⃣1️⃣ Grounded SAM (Zero-Shot Segmentation)")
         with gr.Row():
             seq_gsam_img = gr.Image(label="Grounded SAM Overlay", interactive=False)
             seq_gsam_bw = gr.Image(label="Grounded SAM Binary Mask", interactive=False, image_mode="L")
         seq_gsam_stats = gr.Textbox(label="Grounded SAM Stats", interactive=False)
 
         gr.Markdown("---")
-        gr.Markdown("### 1️⃣1️⃣ Intelliarts Car Parts (Detectron2)")
+        gr.Markdown("### 1️⃣2️⃣ Intelliarts Car Parts (Detectron2)")
         with gr.Row():
             seq_intell_img = gr.Image(label="Intelliarts Car Parts Overlay", interactive=False)
             seq_intell_bw = gr.Image(label="Intelliarts Car Parts Binary Mask", interactive=False, image_mode="L")
@@ -694,63 +766,68 @@ with gr.Blocks(theme=theme, title="Car Window Segmentation") as demo:
 
         def run_all_models(img):
             if img is None: 
-                yield tuple([None]*33)
+                yield tuple([None]*36)
                 return
             
-            # Initialize empty array for all 33 outputs
-            results = [None] * 33
+            # Initialize empty array for all 36 outputs
+            results = [None] * 36
             
-            # 1. YOLOv8x-seg
-            results[0], results[1], results[2] = process_image(img, "YOLOv8x-seg", "", False)
+            # 1. SegFormer
+            results[0], results[1], results[2] = run_segformer(img, morph_cleanup=False)
             yield tuple(results)
             
             # 2. YOLO11x-seg
             results[3], results[4], results[5] = process_image(img, "YOLO11x-seg", "", False)
             yield tuple(results)
             
-            # 3. Mask R-CNN
-            results[6], results[7], results[8] = process_image(img, "Mask R-CNN", "", False)
+            # 3. BiRefNet
+            results[6], results[7], results[8] = run_birefnet(img)
+            yield tuple(results)
+            
+            # 4. YOLOv8x-seg
+            results[9], results[10], results[11] = process_image(img, "YOLOv8x-seg", "", False)
             yield tuple(results)
             
-            # 4. SegFormer
-            results[9], results[10], results[11] = run_segformer(img, morph_cleanup=False)
+            # 5. Mask R-CNN
+            results[12], results[13], results[14] = process_image(img, "Mask R-CNN", "", False)
             yield tuple(results)
             
-            # 5. SegFormer + Morphology
-            results[12], results[13], results[14] = run_segformer(img, morph_cleanup=True)
+            # 6. SegFormer + Morphology
+            results[15], results[16], results[17] = run_segformer(img, morph_cleanup=True)
             yield tuple(results)
             
-            # 6. SAM + YOLO Strat 1
-            results[15], results[16], results[17] = process_image(img, "SAM + YOLO (Strategy 1: Bbox + 5 Points)", "", False)
+            # 7. SAM + YOLO Strat 1
+            results[18], results[19], results[20] = process_image(img, "SAM + YOLO (Strategy 1: Bbox + 5 Points)", "", False)
             yield tuple(results)
             
-            # 7. SAM + YOLO Strat 2
-            results[18], results[19], results[20] = process_image(img, "SAM + YOLO (Strategy 2: Mask + 5 Points)", "", False)
+            # 8. SAM + YOLO Strat 2
+            results[21], results[22], results[23] = process_image(img, "SAM + YOLO (Strategy 2: Mask + 5 Points)", "", False)
             yield tuple(results)
             
-            # 8. SAM + YOLO Strat 3
-            results[21], results[22], results[23] = process_image(img, "SAM + YOLO (Strategy 3: Direct Mask Prompting)", "", False)
+            # 9. SAM + YOLO Strat 3
+            results[24], results[25], results[26] = process_image(img, "SAM + YOLO (Strategy 3: Direct Mask Prompting)", "", False)
             yield tuple(results)
             
-            # 9. Grounding DINO
-            results[24], results[25], results[26] = process_image(img, "Grounding DINO (Zero-Shot Detection)", "car window. car glass. windshield.", False)
+            # 10. Grounding DINO
+            results[27], results[28], results[29] = process_image(img, "Grounding DINO (Zero-Shot Detection)", "car window. car glass. windshield.", False)
             yield tuple(results)
             
-            # 10. Grounded SAM
-            results[27], results[28], results[29] = process_image(img, "Grounded SAM (Zero-Shot Segmentation)", "car window. car glass. windshield.", False)
+            # 11. Grounded SAM
+            results[30], results[31], results[32] = process_image(img, "Grounded SAM (Zero-Shot Segmentation)", "car window. car glass. windshield.", False)
             yield tuple(results)
             
-            # 11. Intelliarts
-            results[30], results[31], results[32] = process_image(img, "Intelliarts Car Parts (Detectron2)", "", False)
+            # 12. Intelliarts
+            results[33], results[34], results[35] = process_image(img, "Intelliarts Car Parts (Detectron2)", "", False)
             yield tuple(results)
 
         submit_btn_seq.click(
             fn=run_all_models,
             inputs=[input_image_seq],
-            outputs=[seq_yolo_img, seq_yolo_bw, seq_yolo_stats, 
+            outputs=[seq_segf_img, seq_segf_bw, seq_segf_stats,
                      seq_yolo11_img, seq_yolo11_bw, seq_yolo11_stats,
+                     seq_biref_img, seq_biref_bw, seq_biref_stats,
+                     seq_yolo_img, seq_yolo_bw, seq_yolo_stats, 
                      seq_mrcnn_img, seq_mrcnn_bw, seq_mrcnn_stats, 
-                     seq_segf_img, seq_segf_bw, seq_segf_stats,
                      seq_segf_morph_img, seq_segf_morph_bw, seq_segf_morph_stats,
                      seq_sam1_img, seq_sam1_bw, seq_sam1_stats,
                      seq_sam2_img, seq_sam2_bw, seq_sam2_stats,