support CMU Mocap

.gitignore

@@ -13,3 +13,4 @@ smplpytorch/native/models/*.pkl
 
 exp/
 output/
+make_gif.py

README.md

@@ -43,6 +43,7 @@ The SMPL human body layer for Pytorch is from the [smplpytorch](https://github.c
     currently supported datasets:
 
     - [HumanAct12](https://ericguo5513.github.io/action-to-motion/)
+    - [CMU Mocap](https://ericguo5513.github.io/action-to-motion/)
     - [UTD-MHAD](https://personal.utdallas.edu/~kehtar/UTD-MHAD.html)
     
 - Set the **DATASET.PATH** in the corresponding configuration file to the location of dataset.

fit/configs/CMU_Mocap.json

@@ -0,0 +1,89 @@
+{
+    "MODEL": {
+        "GENDER": "neutral"
+    },
+    "TRAIN": {
+        "LEARNING_RATE": 2e-2,
+        "MAX_EPOCH": 1000,
+        "WRITE": 1,
+        "OPTIMIZE_SCALE":1,
+        "OPTIMIZE_SHAPE":0
+    },
+    "USE_GPU": 1,
+    "DATASET": {
+        "NAME": "UTD-MHAD",
+        "PATH": "../Action2Motion/CMU Mocap/mocap/mocap_3djoints/",
+        "TARGET_PATH": "",
+        "DATA_MAP": [
+            [
+                0,
+                0
+            ],
+            [
+                9,
+                1
+            ],
+            [
+                12,
+                2
+            ],
+            [
+                13,
+                4
+            ],
+            [
+                18,
+                6
+            ],
+            [
+                20,
+                7
+            ],
+            [
+                14,
+                8
+            ],
+            [
+                19,
+                10
+            ],
+            [
+                21,
+                11
+            ],
+            [
+                1,
+                12
+            ],
+            [
+                4,
+                13
+            ],
+            [
+                7,
+                14
+            ],
+            [
+                10,
+                15
+            ],
+            [
+                2,
+                16
+            ],
+            [
+                5,
+                17
+            ],
+            [
+                8,
+                18
+            ],
+            [
+                11,
+                19
+            ]
+        ]
+    },
+    "DEBUG": 0
+}

fit/configs/Human3.6M.json

@@ -0,0 +1,83 @@
+{
+    "MODEL": {
+        "GENDER": "neutral"
+    },
+    "TRAIN": {
+        "LEARNING_RATE": 2e-2,
+        "MAX_EPOCH": 1000,
+        "WRITE": 1
+    },
+    "USE_GPU": 1,
+    "DATASET": {
+        "NAME": "UTD-MHAD",
+        "PATH": "../UTD-MHAD/Skeleton/Skeleton/",
+        "TARGET_PATH": "",
+        "DATA_MAP": [
+            [
+                12,
+                1
+            ],
+            [
+                0,
+                3
+            ],
+            [
+                16,
+                4
+            ],
+            [
+                18,
+                5
+            ],
+            [
+                20,
+                6
+            ],
+            [
+                22,
+                7
+            ],
+            [
+                17,
+                8
+            ],
+            [
+                19,
+                9
+            ],
+            [
+                21,
+                10
+            ],
+            [
+                23,
+                11
+            ],
+            [
+                1,
+                12
+            ],
+            [
+                4,
+                13
+            ],
+            [
+                7,
+                14
+            ],
+            [
+                2,
+                16
+            ],
+            [
+                5,
+                17
+            ],
+            [
+                8,
+                18
+            ]
+        ]
+    },
+    "DEBUG": 0
+}

fit/configs/HumanAct12.json

@@ -5,7 +5,9 @@
     "TRAIN": {
         "LEARNING_RATE": 2e-2,
         "MAX_EPOCH": 1000,
-        "WRITE": 1
+        "WRITE": 1,
+        "OPTIMIZE_SCALE":0,
+        "OPTIMIZE_SHAPE":1
     },
     "USE_GPU": 1,
     "DATASET": {

fit/configs/UTD_MHAD.json

@@ -5,7 +5,9 @@
     "TRAIN": {
         "LEARNING_RATE": 2e-2,
         "MAX_EPOCH": 1000,
-        "WRITE": 1
+        "WRITE": 1,
+        "OPTIMIZE_SCALE":0,
+        "OPTIMIZE_SHAPE":1
     },
     "USE_GPU": 1,
     "DATASET": {

fit/tools/load.py

@@ -14,3 +14,6 @@ def load(name, path):
         return new_arr
     elif name == 'HumanAct12':
         return np.load(path,allow_pickle=True)
+    elif name == "CMU_Mocap":
+        return np.load(path,allow_pickle=True)
+        

fit/tools/save.py

@@ -33,17 +33,17 @@ def save_pic(res, smpl_layer, file, logger, dataset_name,target):
             savepath=os.path.join(fit_path+"/frame_{}".format(i)),
             batch_idx=i,
             show=False,
-            only_joint=False)
-        display_model(
-            {'verts': verts.cpu().detach(),
-             'joints': target.cpu().detach()},
-            model_faces=smpl_layer.th_faces,
-            with_joints=True,
-            kintree_table=smpl_layer.kintree_table,
-            savepath=os.path.join(gt_path+"/frame_{}".format(i)),
-            batch_idx=i,
-            show=False,
             only_joint=True)
+        # display_model(
+        #     {'verts': verts.cpu().detach(),
+        #      'joints': target.cpu().detach()},
+        #     model_faces=smpl_layer.th_faces,
+        #     with_joints=True,
+        #     kintree_table=smpl_layer.kintree_table,
+        #     savepath=os.path.join(gt_path+"/frame_{}".format(i)),
+        #     batch_idx=i,
+        #     show=False,
+        #     only_joint=True)
     logger.info('Pictures saved')
 
 

fit/tools/train.py

@@ -42,10 +42,10 @@ def init(smpl_layer, target, device, cfg):
     params["scale"] = params["scale"].to(device)
     
     params["pose_params"].requires_grad = True
-    params["shape_params"].requires_grad = True
+    params["shape_params"].requires_grad = bool(cfg.TRAIN.OPTIMIZE_SHAPE)
-    params["scale"].requires_grad = False
+    params["scale"].requires_grad = bool(cfg.TRAIN.OPTIMIZE_SCALE)
     
-    optimizer = optim.Adam([params["pose_params"], params["shape_params"]],
+    optimizer = optim.Adam([params["pose_params"], params["shape_params"], params["scale"]],
                            lr=cfg.TRAIN.LEARNING_RATE)
     
     index={}
@@ -73,8 +73,9 @@ def train(smpl_layer, target,
     
     early_stop = Early_Stop()
     for epoch in tqdm(range(cfg.TRAIN.MAX_EPOCH)):
+    # for epoch in range(cfg.TRAIN.MAX_EPOCH):
         verts, Jtr = smpl_layer(pose_params, th_betas=shape_params)
-        loss = F.smooth_l1_loss(Jtr.index_select(1, index["smpl_index"]) * 100,
+        loss = F.smooth_l1_loss(Jtr.index_select(1, index["smpl_index"]) * 100 * scale,
                                 target.index_select(1, index["dataset_index"]) * 100)
         optimizer.zero_grad()
         loss.backward()
@@ -88,8 +89,8 @@ def train(smpl_layer, target,
             break
         
         if epoch % cfg.TRAIN.WRITE == 0:
-            # logger.info("Epoch {}, lossPerBatch={:.6f}, EarlyStopSatis: {}".format(
+            # logger.info("Epoch {}, lossPerBatch={:.6f}, scale={:.4f} EarlyStopSatis: {}".format(
-            #         epoch, float(loss), early_stop.satis_num))
+            #         epoch, float(loss),float(scale), early_stop.satis_num))
             writer.add_scalar('loss', float(loss), epoch)
             writer.add_scalar('learning_rate', float(
                 optimizer.state_dict()['param_groups'][0]['lr']), epoch)

fit/tools/transform.py

@@ -1,23 +1,17 @@
 import numpy as np
 
+rotate = {
+    'HumanAct12': [1., -1., -1.],
+    'CMU_Mocap': [0.05, 0.05, 0.05],
+    'UTD_MHAD': [-1., 1., -1.]
+}
+
 
 def transform(name, arr: np.ndarray):
-    if name == 'HumanAct12':
-        rotate = [1., -1., -1.]
     for i in range(arr.shape[0]):
         origin = arr[i][0].copy()
         for j in range(arr.shape[1]):
             arr[i][j] -= origin
             for k in range(3):
-                    arr[i][j][k] *= rotate[k]
+                arr[i][j][k] *= rotate[name][k]
-            arr[i][0] = [0.0, 0.0, 0.0]
-    elif name == 'UTD_MHAD':
-        rotate = [-1., 1.,-1.]
-        for i in range(arr.shape[0]):
-            origin = arr[i][3].copy()
-            for j in range(arr.shape[1]):
-                arr[i][j] -= origin
-                for k in range(3):
-                    arr[i][j][k] *= rotate[k]
-            arr[i][3] = [0.0, 0.0, 0.0]
     return arr

make_gif.py

@@ -1,7 +1,7 @@
 import matplotlib.pyplot as plt
 import imageio, os
 images = []
-filenames = sorted(fn for fn in os.listdir('./fit/output/HumanAct12/picture/fit/P01G01R01F0001T0064A0101') )
+filenames = sorted(fn for fn in os.listdir('./fit/output/CMU_Mocap/picture/fit/01_01') )
 for filename in filenames:
-    images.append(imageio.imread('./fit/output/HumanAct12/picture/fit/P01G01R01F0001T0064A0101/'+filename))
+    images.append(imageio.imread('./fit/output/CMU_Mocap/picture/fit/01_01/'+filename))
-imageio.mimsave('./assets/fit.gif', images, duration=0.25)
+imageio.mimsave('fit.gif', images, duration=0.2)