New clip option

Author: semjonsona

src/common_constants.py

@@ -29,6 +29,7 @@ def __init__(self, default_value=None, *args):
     DO_OUTPUT_DEPTH_PREDICTION = False  # Hidden, do not use, subject to change
 
     CLIPDEPTH = False
+    CLIPDEPTH_MODE = "Range"
     CLIPDEPTH_FAR = 0.0
     CLIPDEPTH_NEAR = 1.0
 

src/common_ui.py

@@ -66,6 +66,8 @@ def main_ui_panel(is_depth_tab):
             inp -= 'depthmap_gen_row_3', cur_option_root
             with gr.Row():
                 inp += go.CLIPDEPTH, gr.Checkbox(label="Clip and renormalize DepthMap")
+                inp += go.CLIPDEPTH_MODE,\
+                    gr.Dropdown(label="Mode", choices=['Range', 'Outliers'], type="value", visible=False)
             with gr.Row(visible=False) as clip_options_row_1:
                 inp += go.CLIPDEPTH_FAR, gr.Slider(minimum=0, maximum=1, step=0.001, label='Far clip')
                 inp += go.CLIPDEPTH_NEAR, gr.Slider(minimum=0, maximum=1, step=0.001, label='Near clip')
@@ -179,6 +181,7 @@ def update_default_net_size(model_type):
         inp.add_rule(options_depend_on_output_depth_1, 'visible-if', go.DO_OUTPUT_DEPTH)
         inp.add_rule(go.OUTPUT_DEPTH_INVERT, 'visible-if', go.DO_OUTPUT_DEPTH)
         inp.add_rule(go.OUTPUT_DEPTH_COMBINE_AXIS, 'visible-if', go.OUTPUT_DEPTH_COMBINE)
+        inp.add_rule(go.CLIPDEPTH_MODE, 'visible-if', go.CLIPDEPTH)
         inp.add_rule(clip_options_row_1, 'visible-if', go.CLIPDEPTH)
 
         inp[go.CLIPDEPTH_FAR].change(

src/core.py

@@ -190,8 +190,12 @@ def core_generation_funnel(outpath, inputimages, inputdepthmaps, inputnames, inp
                     if inp[go.DO_OUTPUT_DEPTH_PREDICTION]:
                         yield count, 'depth_prediction', np.copy(out)
                     if inp[go.CLIPDEPTH]:
-                        out = (out - out.min()) / (out.max() - out.min())  # normalize to [0; 1]
-                        out = np.clip(out, inp[go.CLIPDEPTH_FAR], inp[go.CLIPDEPTH_NEAR])
+                        if inp[go.CLIPDEPTH_MODE] == 'Range':
+                            out = (out - out.min()) / (out.max() - out.min())  # normalize to [0; 1]
+                            out = np.clip(out, inp[go.CLIPDEPTH_FAR], inp[go.CLIPDEPTH_NEAR])
+                        elif inp[go.CLIPDEPTH_MODE] == 'Outliers':
+                            fb, nb = np.percentile(out, [inp[go.CLIPDEPTH_FAR] * 100.0, inp[go.CLIPDEPTH_NEAR] * 100.0])
+                            out = np.clip(out, fb, nb)
                     out = (out - out.min()) / (out.max() - out.min())  # normalize to [0; 1]
                 else:
                     # Regretfully, the depthmap is broken and will be replaced with a black image