Pointmap: add per-step timing prints to pin down the bottleneck

app.py

@@ -205,32 +205,44 @@ def _make_ply(image_pil_native: Image.Image, pointmap_hwc: np.ndarray,
 # -----------------------------------------------------------------------------
 # Gradio handler
 
+import time as _t
+
 @spaces.GPU(duration=120)
 def predict(image: Image.Image, size: str):
     if image is None:
         return None, None
 
+    t0 = _t.perf_counter()
     image_pil = image.convert("RGB")
     image_bgr = cv2.cvtColor(np.array(image_pil), cv2.COLOR_RGB2BGR)
+    print(f"[time] convert+bgr           {(_t.perf_counter()-t0)*1000:.0f} ms  (input {image_pil.size})")
 
+    t = _t.perf_counter()
     model = _get_pointmap_model(size)
-    pointmap = _estimate_pointmap(image_bgr, model)        # (H_n, W_n, 3) at most 1024 in either dim
+    print(f"[time] _get_pointmap_model   {(_t.perf_counter()-t)*1000:.0f} ms")
+
+    t = _t.perf_counter()
+    pointmap = _estimate_pointmap(image_bgr, model)
     h_n, w_n = pointmap.shape[:2]
+    print(f"[time] _estimate_pointmap    {(_t.perf_counter()-t)*1000:.0f} ms  (native {w_n}x{h_n})")
 
-    mask = _foreground_mask(image_pil, h_n, w_n)            # native-res mask, fast
+    t = _t.perf_counter()
+    mask = _foreground_mask(image_pil, h_n, w_n)
+    print(f"[time] _foreground_mask      {(_t.perf_counter()-t)*1000:.0f} ms")
 
-    # Depth heatmap (right pane). Solid mid-grey background with the foreground
-    # turbo-coloured by inverse depth. Mirrors sapiens2 vis_pointmap.py colormap.
+    t = _t.perf_counter()
     depth = pointmap[:, :, 2]
     depth_rgb = _depth_to_rgb(depth, mask)
-    BG_GREY = 200
-    depth_rgb[~mask] = BG_GREY
+    depth_rgb[~mask] = 200
     w0, h0 = image_pil.size
     depth_pil = Image.fromarray(depth_rgb).resize((w0, h0), Image.LANCZOS)
+    print(f"[time] depth heatmap+resize  {(_t.perf_counter()-t)*1000:.0f} ms  (target {w0}x{h0})")
 
-    # PLY (download in accordion). Native-res, ≤200K points.
+    t = _t.perf_counter()
     ply_path = _make_ply(image_pil, pointmap, mask)
+    print(f"[time] _make_ply             {(_t.perf_counter()-t)*1000:.0f} ms")
 
+    print(f"[time] TOTAL                 {(_t.perf_counter()-t0)*1000:.0f} ms")
     return depth_pil, ply_path