clean pcd: switch to size-based method

handcrafted_solution.py

@@ -253,7 +253,7 @@ def get_smooth_uv_depth(vertices, depth, gest_seg_np, sfm_depth_np, r=5):
     vertex_depth = np.array(vertex_depth)
     return uv, vertex_depth
 
-
+'''
 from numba import njit, prange
 @njit(parallel=True)
 def fill_range(u, v, z, dilate_r, c, sfm_depth_np, sfm_color_np, H, W):
@@ -267,7 +267,7 @@ def fill_range(u, v, z, dilate_r, c, sfm_depth_np, sfm_color_np, H, W):
             if DUMP_IMG: 
               sfm_color_np[j, i] = c
   return sfm_depth_np, sfm_color_np
-''''''
+'''
 
 def get_SfM_depth(XYZ, rgb, depth_np, gest_seg_np, K, R, t, dilate_r = 5):
   '''Project 3D sfm pointcloud to the image plane '''
@@ -296,7 +296,7 @@ def get_SfM_depth(XYZ, rgb, depth_np, gest_seg_np, K, R, t, dilate_r = 5):
   #checked = 0  
   #print('dim of us uv zs rgb:', len(us), len(vs), len(zs), len(rgb))
   for u,v,z,c in zip(us,vs,zs, rgb):
-    
+    '''
     sfm_depth_np, sfm_color_np = fill_range(u, v, z, dilate_r, c, sfm_depth_np, sfm_color_np, H, W)
     '''
     i_range = range(max(0, u - dilate_r), min(W, u + dilate_r))
@@ -310,7 +310,7 @@ def get_SfM_depth(XYZ, rgb, depth_np, gest_seg_np, K, R, t, dilate_r = 5):
             sfm_depth_np[j, i] = z
             if DUMP_IMG: 
               sfm_color_np[j, i] = c
-    '''
+    
          
               
   #print(f'checked {checked} pts')
@@ -780,14 +780,28 @@ def clean_PCD(XYZ, rgb):
   center_thr = 500
   retain_class_mask = labels == -2 
 
+  largest_blob_size = 0
+  largest_blob = 0
   for k in unique_labels:
-    if k == -1:
-      continue
     class_member_mask = labels == k
+    blob_size = np.count_nonzero(class_member_mask)
+    if blob_size>largest_blob_size:
+      largest_blob_size = blob_size
+      largest_blob = k
+
+  for k in unique_labels:
+    class_member_mask = labels == k
+    if k == -1:
+      retain_class_mask = retain_class_mask | class_member_mask
+      continue    
+    ''' center prior is not valid    
     pt_k = XYZ[class_member_mask]
     Xmean = np.mean(pt_k[:,0])
     Ymean = np.mean(pt_k[:,1])
-    if Xmean < center_thr and Ymean < center_thr:
+    if abs(Xmean) < center_thr and abs(Ymean) < center_thr:
+      retain_class_mask = retain_class_mask | class_member_mask
+    '''
+    if k == largest_blob:
       retain_class_mask = retain_class_mask | class_member_mask
   
   XYZ = XYZ[retain_class_mask]
@@ -799,7 +813,7 @@ def predict(entry, visualize=False, prune_dist_thr=600, depth_scale=2.5, ) -> Tu
     points3D = good_entry['points3d']
     XYZ = np.stack([p.xyz for p in points3D.values()])
     rgb = np.stack([p.rgb for p in points3D.values()])
-    #XYZ, rgb = clean_PCD(XYZ, rgb)
+    XYZ, rgb = clean_PCD(XYZ, rgb)
     vert_edge_per_image = {}
     for i, (ade, gest, depth, K, R, t) in enumerate(zip(
                                                 good_entry['ade20k'],
@@ -860,7 +874,9 @@ def predict(entry, visualize=False, prune_dist_thr=600, depth_scale=2.5, ) -> Tu
     #print('before merge, vertices=', vertices, ' connections=', connections, ' vertices_3d=', vertices_3d)
     #all_3d_vertices, connections_3d = merge_vertices_3d(vert_edge_per_image, 3.0) # TODO: 3cm looks too small       
     all_3d_vertices, connections_3d = merge_vertices_3d(vert_edge_per_image, 150)
+    #all_3d_vertices, connections_3d = merge_vertices_3d(vert_edge_per_image, 50)
     #print(f'after merge, {len(all_3d_vertices)} 3d vertices and {len(connections_3d)} 3d connections')
+    #print(f'after merge, 3d vertices: {all_3d_vertices} and 3d connections: {connections_3d}')
     #all_3d_vertices_clean, connections_3d_clean  = prune_not_connected(all_3d_vertices, connections_3d)    
     all_3d_vertices, connections_3d  = prune_tall_short(all_3d_vertices, connections_3d, 1000, 0)