WIP: scale keypoints to match

camera_estimation.py

@@ -124,7 +124,7 @@ class CameraEstimate:
             init_points_3d = torch.from_numpy(init_points_3d)
 
             params = [translation, rotation]
-            opt = torch.optim.Adam(params, lr=0.1)
+            opt = torch.optim.Adam(params, lr=0.001)
 
             def C(params, X):
                 translation = params[0]
@@ -146,7 +146,7 @@ class CameraEstimate:
                 loss.float()
                 loss.backward()
                 opt.step()
-                stop = loss > 3e-4
+                stop = loss > 3e-2
                 current_pose = self.torch_params_to_pose(params)
                 current_pose = current_pose.detach().numpy()
                 self.renderer.scene.set_pose(self.transformed_points, current_pose)

example_fitter.py

@@ -12,7 +12,7 @@ from model import *
 import time
 # from renderer import *
 from dataset import *
-from utils import get_named_joints, estimate_depth
+from utils import get_named_joints, estimate_scale
 
 ascii_logo = """\
   /$$$$$$  /$$      /$$ /$$$$$$$  /$$   /$$     /$$
@@ -66,10 +66,10 @@ joints = model_out.joints.detach().cpu().numpy().squeeze()
 # ---------------------------------
 cam_est_joints_names = ["hip-left", "hip-right",
                         "shoulder-left", "shoulder-right"]
-est_depth = estimate_depth(joints, keypoints)
+est_scale = estimate_scale(joints, keypoints)
 
-# apply depth to keypoints
-keypoints[:, 2] = -est_depth
+# apply scaling to keypoints
+keypoints = keypoints * est_scale
 
 init_joints = get_named_joints(joints, cam_est_joints_names)
 init_keypoints = get_named_joints(keypoints, cam_est_joints_names)
@@ -108,7 +108,7 @@ keyp_torso = torch.Tensor(init_keypoints, device=device)
 
 learning_rate = 1e-3
 trans = Transform(dtype, device)
-proj = CameraProjSimple(dtype, device, -est_depth)
+proj = CameraProjSimple(dtype, device, 1)
 optimizer = torch.optim.Adam(trans.parameters(), lr=learning_rate)
 loss_layer = torch.nn.MSELoss()
 

modules/transform.py

@@ -66,5 +66,5 @@ class Transform(nn.Module):
         return transform
 
     def forward(self, joints):
-        R = self.get_transform_mat().squeeze()
+        R = self.get_transform_mat()
         return joints @ R + F.pad(self.translation, (0,1), value=1)

renderer.py

@@ -18,8 +18,6 @@ class Renderer:
         if camera is None:
             camera = pyrender.OrthographicCamera(ymag=1, xmag=1)
 
-       
-
         if camera_pose is None:
             camera_pose = np.eye(4)
             camera_pose[:3, :3] = R.from_rotvec(np.pi/2 * np.array([0, 0, 0])).as_matrix()

utils.py

@@ -130,8 +130,8 @@ def openpose_to_opengl_coords(
 
     points = np.array([
         [
-            x / real_width * 2 - 1,
-            -y / real_height * 2 + 1,
+            x / real_width * 5 - 1,
+            -y / real_height * 5 + 1,
             0
         ] for (x, y, z) in input_data])
 
@@ -176,11 +176,11 @@ def render_points(scene, points, radius=0.005, color=[0.0, 0.0, 1.0, 1.0], name=
     tfs = np.tile(np.eye(4), (len(points), 1, 1))
     tfs[:, :3, 3] = points
     pcl = pyrender.Mesh.from_trimesh(sm, poses=tfs)
-    # return the render scene node
+    # return the render scsene node
     return scene.add(pcl, name=name)
 
 
-def estimate_depth(joints, keypoints, pairs=[
+def estimate_scale(joints, keypoints, pairs=[
     ("shoulder-right", "hip-right"),
     ("shoulder-left", "hip-left")
 ], cam_fy=1):