check flame config

apps/calibration/camera_parameters.md

@@ -1,19 +0,0 @@
-<!--
- * @Date: 2021-04-13 16:49:12
- * @Author: Qing Shuai
- * @LastEditors: Qing Shuai
- * @LastEditTime: 2021-04-13 16:51:16
- * @FilePath: /EasyMocapRelease/apps/calibration/camera_parameters.md
--->
-# Camera Parameters Format
-
-For example, if the name of a video is `1.mp4`, then there must exist `K_1`, `dist_1` in `intri.yml`, and `R_1((3, 1), rotation vector of camera)`, `T_1(3, 1)` in `extri.yml`. The file format is following [OpenCV format](https://docs.opencv.org/master/dd/d74/tutorial_file_input_output_with_xml_yml.html).
-
-## Write/Read
-
-See `easymocap/mytools/camera_utils.py`=>`write_camera`, `read_camera` functions.
-
-## Conversion between different format
-
-TODO
-

apps/mocap/run.py

@@ -3,14 +3,27 @@ import os
 from easymocap.config import Config, load_object
 from tqdm import tqdm
 
-def process(dataset, model):
+def process(dataset, model, args):
     ret_all = []
     print('[Run] dataset has {} samples'.format(len(dataset)))
-    for i in tqdm(range(len(dataset)), desc='[Run]'):
+    if args.num_workers == -1:
-        data = dataset[i]
+        for i in tqdm(range(len(dataset)), desc='[Run]'):
-        ret = model.at_step(data, i)
+            data = dataset[i]
-        ret_all.append(ret)
+            ret = model.at_step(data, i)
-    ret_all = model.at_final(ret_all)
+            ret_all.append(ret)
+    else:
+        import torch
+        dataloader = torch.utils.data.DataLoader(dataset, 
+            batch_size=1, num_workers=args.num_workers, shuffle=False, collate_fn=lambda x:x, drop_last=False)
+        index = 0
+        for data in tqdm(dataloader, desc='[Run]'):
+            data = data[0]
+            ret = model.at_step(data, index)
+            if not args.skip_final:
+                ret_all.append(ret)
+            index += 1
+    if not args.skip_final:
+        ret_all = model.at_final(ret_all)
 
 def update_data_by_args(cfg_data, args):
     if args.root is not None:
@@ -23,7 +36,7 @@ def update_data_by_args(cfg_data, args):
         cfg_data.args.ranges = args.ranges
     if args.cameras is not None:
         cfg_data.args.reader.cameras.root = args.cameras
-    if args.skip_vis:
+    if args.skip_vis or args.skip_vis_step:
         cfg_data.args.subs_vis = []
     return cfg_data
 
@@ -41,7 +54,7 @@ def update_exp_by_args(cfg_exp, args):
                 val.skip = True
     if args.skip_vis or args.skip_vis_final:
         for key, val in cfg_exp.args.at_final.items():
-            if key.startswith('vis'):
+            if key.startswith('vis') or key == 'make_video':
                 val.skip = True    
 
 def load_cfg_from_file(cfg, args):
@@ -74,9 +87,11 @@ def main_entrypoint():
     parser.add_argument('--ranges', type=int, default=None, nargs=3)
     parser.add_argument('--cameras', type=str, default=None, help='Camera file path')
     parser.add_argument('--out', type=str, default=None)
+    parser.add_argument('--num_workers', type=int, default=-1)
     parser.add_argument('--skip_vis', action='store_true')
     parser.add_argument('--skip_vis_step', action='store_true')
     parser.add_argument('--skip_vis_final', action='store_true')
+    parser.add_argument('--skip_final', action='store_true')
     parser.add_argument('--debug', action='store_true')
     args = parser.parse_args()
 
@@ -96,7 +111,7 @@ def main_entrypoint():
     print(dataset)
 
     model = load_object(cfg_exp.module, cfg_exp.args)
-    process(dataset, model)
+    process(dataset, model, args)
 
 if __name__ == '__main__':
     main_entrypoint()

config/model/flame.yml

@@ -0,0 +1,4 @@
+module: easymocap.bodymodel.smpl.SMPLModel
+args:
+  model_path: data/bodymodels/FLAME2020/FLAME_NEUTRAL.pkl
+  device: cuda

easymocap/bodymodel/smpl.py

@@ -85,12 +85,13 @@ class SMPLModel(Model):
         self.lbs = lbs
         self.data = load_model_data(model_path)
         self.register_any_lbs(self.data)
-
         # keypoints regressor
         if regressor_path is not None and use_joints:
             X_regressor = load_regressor(regressor_path)
             X_regressor = torch.cat((self.J_regressor, X_regressor), dim=0)
             self.register_any_keypoints(X_regressor)
+        elif regressor_path is None:
+            self.register_any_keypoints(self.J_regressor)
         if not self.use_root_rot:
             self.NUM_POSES -= 3 # remove first 3 dims
         self.to(self.device)
@@ -169,7 +170,9 @@ class SMPLModel(Model):
 
     def forward(self, return_verts=True, return_tensor=True, 
         return_smpl_joints=False, 
-        only_shape=False, pose2rot=True, **params):
+        only_shape=False, pose2rot=True,
+        v_template=None,
+        **params):
         params = self.check_params(params)
         poses, shapes = params['poses'], params['shapes']
         poses = self.extend_poses(pose2rot=pose2rot, **params)
@@ -187,7 +190,8 @@ class SMPLModel(Model):
             Rh = batch_rodrigues(Rh)
         Th = Th.unsqueeze(dim=1)
         if return_verts or not self.use_joints:
-            v_template = self.v_template
+            if v_template is None:
+                v_template = self.v_template
             if 'scale' in params.keys():
                 v_template = v_template * params['scale'][0]
             vertices, joints, T_joints, T_vertices = self.lbs(shapes, poses, v_template,
@@ -199,7 +203,8 @@ class SMPLModel(Model):
                 vertices = joints
         else:
             # only forward joints
-            v_template = self.j_v_template
+            if v_template is None:
+                v_template = self.j_v_template
             if 'scale' in params.keys():
                 v_template = v_template * params['scale'][0]
             vertices, joints, _, _ = self.lbs(shapes, poses, v_template,
@@ -236,6 +241,39 @@ class SMPLModel(Model):
         return Params.merge(params, **kwargs)
 
 
+    def convert_to_standard_smpl(self, params):
+        params = self.check_params(params)
+        poses, shapes = params['poses'], params['shapes']
+        Th = params['Th']
+        vertices, joints, _, _ = lbs(shapes, poses, self.v_template,
+                                self.shapedirs, self.posedirs,
+                                self.J_regressor, self.parents,
+                                self.weights, pose2rot=True, dtype=self.dtype, only_shape=True)
+        # N x 3
+        j0 = joints[:, 0, :]
+        Rh = params['Rh']
+        # N x 3 x 3
+        rot = batch_rodrigues(Rh)
+        # change the rotate center
+        min_xyz, _ = self.v_template.min(dim=0, keepdim=True)
+        # X' = X + delta_center
+        delta_center = torch.tensor([0., -min_xyz[0, 1], 0.]).reshape(1, 3).to(j0.device)
+        # J' = J + delta_center
+        j0new = j0 + delta_center
+        # Tnew = T - (R(d - J0) + J0)
+        Tnew = Th - (torch.einsum('bij,bj->bi', rot, delta_center-j0new) + j0new)
+        if poses.shape[1] == 69:
+            poses = torch.cat([Rh, poses], dim=1)
+        elif poses.shape[1] == 72:
+            poses[:, :3] = Rh
+        else:
+            import ipdb;ipdb.set_trace()
+        res = dict(poses=poses.detach().cpu().numpy(),
+            shapes=shapes.detach().cpu().numpy(),
+            Th=Tnew.detach().cpu().numpy()
+            )
+        return res
+    
     def convert_from_standard_smpl(self, params):
         params = self.check_params(params)
         poses, shapes = params['poses'], params['shapes']