eval_pose_fist

.gitignore

@@ -49,4 +49,8 @@ models/**/build
 models/**/dist
 
 temp_local
-examples/results
+examples/results
+
+dyn_check
+
+evaluation/saved

config/magic_infer_offline.yaml

@@ -39,7 +39,7 @@ model:
       window_len: 60
     stablizer: True
     mode: "online"
-    s_wind: 200
+    s_wind: 500
     overlap: 4
   track_num: 0
 

models/SpaTrackV2/models/vggt4track/models/vggt_moe.py

@@ -31,6 +31,8 @@ class VGGT4Track(nn.Module, PyTorchModelHubMixin):
         self,
         images: torch.Tensor,
         annots = {},
+        fx_prev = None,
+        fy_prev = None,
         **kwargs):
         """
         Forward pass of the VGGT4Track model.
@@ -85,16 +87,29 @@ class VGGT4Track(nn.Module, PyTorchModelHubMixin):
                 predictions["unc_metric"] = depth_conf.view(B*T, H_proc, W_proc)
 
         predictions["images"] = (images)*255.0
+                                                         
         # output the camera pose
         predictions["poses_pred"] = torch.eye(4)[None].repeat(T, 1, 1)[None]
         predictions["poses_pred"][:,:,:3,:4], predictions["intrs"] = pose_encoding_to_extri_intri(predictions["pose_enc_list"][-1],
                                                                                                                      images_proc.shape[-2:])
         predictions["poses_pred"] = torch.inverse(predictions["poses_pred"])
+
+        if fx_prev is not None:
+            scale_x = torch.from_numpy(fx_prev).to(predictions["intrs"].device) / predictions["intrs"][0, :fx_prev.shape[0], 0, 0]
+            scale_x = scale_x.mean() * W_proc / W 
+            predictions["intrs"][:, :, 0, 0] *= scale_x
+        if fy_prev is not None:
+            scale_y = torch.from_numpy(fy_prev).to(predictions["intrs"].device) / predictions["intrs"][0, :fy_prev.shape[0], 1, 1]
+            scale_y = scale_y.mean() * H_proc / H
+            predictions["intrs"][:, :, 1, 1] *= scale_y
+
+        # get the points map
         points_map = depth_to_points_colmap(depth.view(B*T, H_proc, W_proc), predictions["intrs"].view(B*T, 3, 3))
         predictions["points_map"] = points_map
         #NOTE: resize back
         predictions["points_map"] = F.interpolate(points_map.permute(0,3,1,2),
                                                          size=(H, W), mode='bilinear', align_corners=True).permute(0,2,3,1)
+
         predictions["unc_metric"] = F.interpolate(predictions["unc_metric"][:,None],
                                                          size=(H, W), mode='bilinear', align_corners=True)[:,0]
         predictions["intrs"][..., :1, :] *= W/W_proc 

models/SpaTrackV2/models/vggt4track/utils/load_fn.py

@@ -145,7 +145,7 @@ def load_and_preprocess_images(image_path_list, mode="crop"):
 
     return images
 
-def preprocess_image(img_tensor, mode="crop", target_size=518):
+def preprocess_image(img_tensor, mode="crop", target_size=518, keep_ratio=False):
     """
     Preprocess image tensor(s) to target size with crop or pad mode.
     Args:
@@ -190,9 +190,10 @@ def preprocess_image(img_tensor, mode="crop", target_size=518):
             new_W = target_size
             new_H = round(H * (new_W / W) / 14) * 14
             out = torch.nn.functional.interpolate(img.unsqueeze(0), size=(new_H, new_W), mode="bicubic", align_corners=False).squeeze(0)
-            if new_H > target_size:
-                start_y = (new_H - target_size) // 2
-                out = out[:, start_y : start_y + target_size, :]
+            if keep_ratio==False:
+                if new_H > target_size:
+                    start_y = (new_H - target_size) // 2
+                    out = out[:, start_y : start_y + target_size, :]
         processed.append(out)
     result = torch.stack(processed)
     if squeeze:

models/moge/train/losses.py

@@ -46,7 +46,6 @@ def affine_invariant_global_loss(
     scale, shift = torch.where(valid, scale, 0), torch.where(valid[..., None], shift, 0)
 
     pred_points = scale[..., None, None, None] * pred_points + shift[..., None, None, :]
-
     # Compute loss
     weight = (valid[..., None, None] & mask).float() / gt_points[..., 2].clamp_min(1e-5)
     weight = weight.clamp_max(10.0 * weighted_mean(weight, mask, dim=(-2, -1), keepdim=True))   # In case your data contains extremely small depth values