Update configs/tracking.yaml

configs/tracking.yaml

@@ -1,9 +1,9 @@
 config_name : test
 
-batch_size : 16
+batch_size : 16 # batch size used during joint optimization stage, when your video has less frames than batch_size the script will crash currently
 num_views : 1
-size : 256
-image_size : [256, 256] # use this instead hardcoding a bunch of 512 and 256
+size : 256 # defines the rendering resolution, you could also try 512
+image_size : [256, 256] # make sure to update this accordingly, sorry for the redundancy
 
 data_folder : none
 p3dmm_folder : none
@@ -16,37 +16,39 @@ start_frame : 0
 num_shape_params : 300
 num_exp_params : 100
 tex_params : 140
-iters : 200 #800 #400
+iters : 800 #400 # Number if iterations in first stage of tracking. For single-image fitting, you might want to choose a larger value, e.g. 500 or 800
+global_iters : 5000 # Number f iterations for second stage of tracking (video only)
 
-no_lm : False
-use_eyebrows : False
-use_mouth_lmk : True
-no_pho : True
-no_sh : True
-disable_edge : False
 
-keyframes : []
+no_lm : False # disable all landmark losses (only eye and mouth landmarks are used)
+use_eyebrows : False # not sure if used
+use_mouth_lmk : True # toggle use of mouth landmarks
+no_pho : True # unused
+no_sh : True # unused
+disable_edge : False # unused
 
-ignore_mica : False
-flame2023 : False
+keyframes : [] # unused
 
-uv_map_super : 2000.0 #500.0 #100 #2001.0 #5000.0 #2000.0
-normal_super : 1000.0 #202.0
-normal_super_can : 0.0
-sil_super : 500
+ignore_mica : False # unused
+flame2023 : False # unused
 
+uv_map_super : 2000.0 #500.0 #100 #2001.0 #5000.0 #2000.0 # weight for UV-loss
+normal_super : 1000.0 #202.0 # weight for normal loss
+normal_super_can : 0.0 # unused
+sil_super : 500 # weight for silhouette loss
 
+# parameters for UV loss
 uv_loss:
   stricter_uv_mask : False
-  delta_uv : 0.00005 #0.00005 #0.0005 #0.00005
-  delta_uv_fine : 0.00005 #0.00005 #0.0005 #0.00005
-  dist_uv : 20 #20 #15
-  dist_uv_fine : 20 #35 #20 #15
+  delta_uv : 0.00005 #0.00005 #0.0005 #0.00005 # threshold that determines if correspondence is valid or not, ref. to eq.7 of the paper
+  delta_uv_fine : 0.00005 #0.00005 #0.0005 #0.00005 # unused
+  dist_uv : 20 #20 #15 # correspondences with a higher distance than this are considered outliers
+  dist_uv_fine : 20 #35 #20 #15 # unused
 
 
 occ_filter : True
 
-
+# learning rates for different optimizable parameters
 lr_id : 0.002 #0.003 #0.006 #0.003
 lr_exp : 0.005 # 0.005 #0.01 #0.01 #0.005
 lr_jaw : 0.005 #0.003
@@ -56,31 +58,33 @@ lr_t : 0.001 #0.002 #0.001 #0.0005 #0.0005
 lr_f : 0.1 #0.05 #0.01 #0.001
 lr_pp : 0.00005
 
-w_pho : 150
-w_lmks : 3000
-w_lmks_mouth : 1000
-w_lmks_68 : 1000
-w_lmks_lid : 1000
-w_lmks_iris : 1000
-w_lmks_oval : 2000
-w_lmks_star : 0
+# some loss weights
+w_pho : 150 # unused
+w_lmks : 3000 # weight for eye landmarks
+w_lmks_mouth : 1000 # weight for mouth landmarks
+w_lmks_68 : 1000 # unused
+w_lmks_lid : 1000 # weightr for eye closure loss
+w_lmks_iris : 1000 # weight for iris/pupil loss, responsible for viewing direction
+w_lmks_oval : 2000 # unused
+w_lmks_star : 0 # unused
 
-include_neck : True
+include_neck : True # whether to optimize for neck or not
 
+# regualization weights for different optimizable parameters
 w_shape :  0.2
 w_shape_general : 0.05
 w_exp : 0.05
 w_jaw : 0.01
 w_neck : 0.1
 
-n_fine : False
-low_overhead : False
+n_fine : False # unused
+low_overhead : False # unused
 
-delta_n : 0.33
+delta_n : 0.33 # threshold that determines above which loss pixels in the normal loss map are considered outliers
 
-global_camera : True
-global_iters : 10 # reducing it to 10 for Huggingface Space 5000
+global_camera : True 
 
+# smoothness loss weights 
 reg_smooth_exp : 50.0
 reg_smooth_eyes : 10.0
 reg_smooth_eyelids : 2.0
@@ -98,13 +102,13 @@ normal_l2 : False
 smooth : True
 normal_mask_ksize : 13
 
-early_stopping_delta : 5.0
+early_stopping_delta : 5.0 # a higher value will make stage 1 tracking faster but less accurate
 
 early_exit : False
 
-draw_uv_corresp : False
+draw_uv_corresp : True # add additional visualizations to result.mp4, e.g. uv-loss and normal loss map
 
 save_landmarks : False
 
 save_meshes : True
-delete_preprocessing : True
+delete_preprocessing : False