20190727 Update toy dataset link

Author: Eric Yang

PREPS.md

@@ -20,6 +20,7 @@ Make sure you have an GPU card with at least 12 GB graphic memory. You can decre
 We hereby datail the necessary steps to reproduce human36m experiments. Most people failed on this step, so I hope this instruction could make sure you don't become one of them.
 
 1. I tentatively suggest users to first try download human36m dataset from the [official website](http://vision.imar.ro/human3.6m/description.php). However manual authorization could take anywhere from 6 days to 6 months, so if you are keen to get things going, you can start by playing with this [toy_dataset](https://pan.baidu.com/s/1szhb9B_8n6p6CeAoPUxnhw). Extraction code: 0o95 
+Alternative link [Google Drive](https://drive.google.com/open?id=1ssuUje20x1PS5qYwbg1AAloDsVwF1eTW)
 
 2. Unpack downloaded zip files into a single folder (which I suggest you name it `human36m`) and put it under `path-to-your-datasets`. 
 Open `data_utils/data_washing.py`, change the `root_dir` variable in main function near line 190 to your h36m dataset path and run it to perform standard data augmentations, the washed dataset will be stored in `path-to-your-datasets/human36m_washed`.

README.md

@@ -5,6 +5,9 @@ PyTorch implementation of CloudWalk's recent paper [DenseBody](https://arxiv.org
 
 **Update on 20190613** A toy dataset has been released to facilitate the reproduction of this project. checkout [`PREPS.md`](PREPS.md) for details.
 
+**Update on 20190727** Update an alternative [link](https://drive.google.com/open?id=1ssuUje20x1PS5qYwbg1AAloDsVwF1eTW) for the toy dataset.
+
+
 ![paper teaser](teaser/teaser.jpg)
 
 ### Reproduction results

data_utils/uv_map_generator.py

@@ -371,7 +371,7 @@ def UV_interp(self, rgbs):
     UV_map: [H * W * 3] Interpolated UV map.
     colored_verts: [H * W * 3] Scatter plot of colorized UV vertices
     '''
-    def get_UV_map(self, verts):
+    def get_UV_map(self, verts, dilate=True):
         # normalize all to [0,1]
         _min = np.amin(verts, axis=0, keepdims=True)
         _max = np.amax(verts, axis=0, keepdims=True)
@@ -383,8 +383,10 @@ def get_UV_map(self, verts):
         ])
         rgbs = verts[vt_to_v_index]
 
-        #return self.UV_interp(rgbs), verts_backup
-        return self._dilate(self.UV_interp(rgbs)), verts_backup
+        uv_map = self.UV_interp(rgbs)
+        if dilate:
+            uv_map = self._dilate(uv_map)
+        return uv_map, verts_backup
 
     '''
         TODO: make it torch.
@@ -419,23 +421,24 @@ def resample(self, UV_map):
 if __name__ == '__main__':
     # test render module
     # change this to the same as in train.py opt.uv_prefix
-    file_prefix = 'radvani_template'
+    # file_prefix = 'radvani_template'
     #file_prefix = 'vbml_close_template'
     #file_prefix = 'vbml_spaced_template'
+    file_prefix = 'smpl_fbx_template'
     generator = UV_Map_Generator(
-        UV_height=256,
+        UV_height=512,
         UV_pickle=file_prefix+'.pickle'
     )
-    test_folder = '_test_radvani'
+    test_folder = 'smpl_512'
     if not os.path.isdir(test_folder):
         os.makedirs(test_folder)
 
     generator.render_UV_atlas('{}/{}_atlas.png'.format(test_folder, file_prefix))
     img, verts, rgbs = generator.render_point_cloud('{}/{}.png'.format(test_folder, file_prefix))
     verts, rgbs = generator.write_ply('{}/{}.ply'.format(test_folder, file_prefix), verts, rgbs)
-    uv, _ = generator.get_UV_map(verts)
+    uv, _ = generator.get_UV_map(verts, dilate=False)
     uv = uv.max(axis=2)
     print(uv.shape)
     binary_mask = np.where(uv > 0, 1., 0.)
     binary_mask = (binary_mask * 255).astype(np.uint8)
-    imsave('{}_UV_mask.png'.format(file_prefix), binary_mask)
+    imsave('./{}_UV_mask.png'.format(file_prefix), binary_mask)

models/networks.py

@@ -118,7 +118,8 @@ def acquire_weights(UV_weight_npy):
         mask_name = UV_weight_npy.replace('weights.npy', 'mask.png')
         print(mask_name)
         UV_mask = imread(mask_name)
-        UV_mask = UV_mask[:,:,0]
+        if UV_mask.ndim == 3:
+            UV_mask = UV_mask[:,:,0]
         ret, labels = connectedComponents(UV_mask, connectivity=4)
         unique, counts = np.unique(labels, return_counts=True)
         print(unique, counts)
@@ -404,5 +405,5 @@ def forward(self, x):
         return self.conv(x)
 
 if __name__ == '__main__':
-    acquire_weights('../data_utils/radvani_UV_weights.npy')
+    acquire_weights('../data_utils/smpl_fbx_template_UV_weights.npy')
 

train.py

@@ -23,10 +23,8 @@ def TrainOptions(debug=False):
 
     parser.add_argument('--data_root', type=str, default=data_root)
     parser.add_argument('--checkpoints_dir', type=str, default='checkpoints')
-    parser.add_argument('--dataset', type=str, default='human36m',
-        choices=['human36m', 'surreal', 'up3d'])
     parser.add_argument('--max_dataset_size', type=int, default=-1)
-    parser.add_argument('--im_size', type=int, default=256)
+    parser.add_argument('--im_size', type=int, default=512)
     parser.add_argument('--batch_size', type=int, default=batch_size)
     parser.add_argument('--name', type=str, default='densebody_resnet_h36m')
     parser.add_argument('--uv_map', type=str, default='radvani', choices=['radvani', 'vbml_close', 'vbml_spaced', 'smpl_fbx'])