test postometro pipeline

common/utils/vis.py

@@ -138,6 +138,20 @@ def perspective_projection(vertices, cam_param):
     vertices[:, 1] = vertices[:, 1] * fy / vertices[:, 2] + cy
     return vertices
 
+def orthographic_projection(X, cam_param):
+    """Perform orthographic projection of 3D points X using the camera parameters
+    Args:
+        X: size = [N, 3]
+        camera: size = [4]: sx, sy, tx, ty
+    Returns:
+        Projected 2D points -- size = [N, 2]
+    """ 
+    sx, sy, tx, ty = cam_param
+    X_trans = X[:,:2].copy()
+    X_trans[:,0] = (X_trans[:,0] + tx) * sx
+    X_trans[:,1] = (X_trans[:,1] + ty) * sy
+    return X_trans.copy()
+
 class WeakPerspectiveCamera(pyrender.Camera):
     def __init__(self, scale, translation, znear=pyrender.camera.DEFAULT_Z_NEAR, zfar=None, name=None):
         super(WeakPerspectiveCamera, self).__init__(znear=znear, zfar=zfar, name=name)
@@ -156,7 +170,11 @@ class WeakPerspectiveCamera(pyrender.Camera):
 def render_mesh(img, mesh, face, cam_param, mesh_as_vertices=False):
     if mesh_as_vertices:
         # to run on cluster where headless pyrender is not supported for A100/V100
-        vertices_2d = perspective_projection(mesh, cam_param)
+        viewport_height, viewport_width = img.shape[0], img.shape[1]
+        vertices_2d = orthographic_projection(mesh, cam_param) # [-1,1]
+        vertices_2d = (vertices_2d + 1) / 2
+        vertices_2d[:,0] *= viewport_width
+        vertices_2d[:,1] *= viewport_height
         img = vis_keypoints(img, vertices_2d, alpha=0.8, radius=2, color=(0, 0, 255))
     else:
         # mesh

main/inference.py

@@ -100,10 +100,10 @@ class Inferer:
             ok_bboxes.append(bbox)
 
             # [DEBUG] test mmdet pipeline
-            # if bbox is not None:
-            #     top_left = (int(bbox[0]), int(bbox[1]))
-            #     bottom_right = (int(bbox[0] + bbox[2]), int(bbox[1] + bbox[3]))
-            #     cv2.rectangle(vis_img, top_left, bottom_right, (0, 0, 255), 2)
+            if bbox is not None:
+                top_left = (int(bbox[0]), int(bbox[1]))
+                bottom_right = (int(bbox[0] + bbox[2]), int(bbox[1] + bbox[3]))
+                cv2.rectangle(vis_img, top_left, bottom_right, (0, 0, 255), 2)
     
             # human model inference
             img, img2bb_trans, bb2img_trans = generate_patch_image(original_img, bbox, 1.0, 0.0, False, self.cfg.input_img_shape)
@@ -141,12 +141,12 @@ class Inferer:
 
 if __name__ == '__main__':
     from PIL import Image
-    inferer = Inferer('postometro', 0, './out_folder') # gpu
+    inferer = Inferer('postometro', 1, './out_folder') # gpu
     image_path = f'../assets/07.jpg'
     image = Image.open(image_path)
     # Convert the PIL image to a NumPy array
     image_np = np.array(image)
-    vis_img, _ , _ = inferer.infer(image_np, 0.2, multi_person=True, mesh_as_vertices=False)
+    vis_img, _ , _ = inferer.infer(image_np, 0.2, multi_person=True, mesh_as_vertices=True)
     save_path = f'./saved_vis_07.jpg'
 
     # Ensure the image is in the correct format (PIL expects uint8)