facial tremor testing

dbm_lib/dbm_features/raw_features/movement/__init__.py

@@ -13,3 +13,4 @@ import os
 DBMLIB_PATH = os.path.dirname(__file__)
 DBMLIB_VTREMOR_LIB = os.path.abspath(os.path.join(DBMLIB_PATH,
                                                   '../../../../resources/libraries/voice_tremor.praat'))
+DBMLIB_FTREMOR_CONFIG = os.path.abspath(os.path.join(DBMLIB_PATH, '../resources/features/facial/config.json'))

dbm_lib/dbm_features/raw_features/movement/facial_tremor.py

@@ -0,0 +1,159 @@
+import sys, os, glob, cv2, re
+import pickle, json
+import pandas as pd
+import numpy as np
+import numpy.ma as ma
+import logging
+from os.path import join
+
+from dbm_lib.dbm_features.raw_features.util import util as ut
+from dbm_lib.dbm_features.raw_features.util.math_util import *
+
+from dbm_lib.dbm_features.raw_features.movement import DBMLIB_FTREMOR_CONFIG
+
+logging.basicConfig(level=logging.INFO)
+logger=logging.getLogger()
+
+ft_dir = 'movement/facial_tremor'
+csv_ext = '_fac_tremor.csv'
+model_ext = '_fac_model.csv'
+fac_features_ext = '_fac_features.csv'
+
+def compute_features(out_dir, df_of, r_config):
+    """ Computes features
+
+    Returns: features in vector format
+    """
+    config = json.loads(open(DBMLIB_FTREMOR_CONFIG,'r').read())
+    logger.info('json file read')
+    pattern_x = re.compile("l\d+_x")
+    pattern_y = re.compile("l\d+_y")
+
+    # assumption: distance of face to camera remains at roughly static
+
+    # logic break
+    landmark_columns = []
+    for col in df_of.columns:
+        if pattern_x.match(col) or pattern_y.match(col):
+            landmark_columns.append(col)
+
+    df_of= df_of[(df_of[landmark_columns]!= 0).any(axis=1)]
+    df_of.reset_index(inplace=True)
+
+    num_frames = len(df)
+    logger.info("Number of frames to be processed: {}".format(str(num_frames)))
+    landmarks = config['landmarks']
+
+    try:
+        if num_frames == 0:
+            error_reason = "No frames with visible face."
+            logger.error(error_reason)
+            return empty_frame(landmarks, r_config, error_reason)
+
+#             if num_frames < 60:
+#                 error_reason = 'Number of frames with visible face < 60. Video too short'
+#                 logger.error(error_reason)
+#                 return empty_frame(landmarks, f_cfg, error_reason)
+
+        first_row = df_of.iloc[0]
+
+        facew = abs(first_row[config['face_width_left']] - first_row[config['face_width_right']])
+        faceh = abs(first_row[config['face_height_left']] - first_row[config['face_height_right']])
+
+        if facew == 0 or faceh == 0:
+            error_reason = 'face width or height = 0. Check landmark values'
+            logger.error(error_reason)
+            return empty_frame(landmarks, r_config)
+
+        fac_disp = calc_displacement_vec(df_of, landmarks, num_frames)
+
+        # if verbose:
+            # logger.info("Displacement output: {}".format(str(fac_disp)))
+
+        fac_disp_median = np.median(fac_disp, axis = 1)
+        fac_disp_mean = np.mean(fac_disp, axis = 1)
+
+        if len(fac_disp.shape)!=2:
+            error_reason = 'fac_disp is not 2D. smth went wrong with disp calc'
+            logger.error(error_reason)
+            return empty_frame(landmarks, r_config, error_reason)
+
+        if len(fac_disp[0])<=1:
+            error_reason = 'Video too short. smth went wrong with disp calc'
+            logger.error(error_reason)
+            return empty_frame(landmarks, r_config, error_reason)
+
+        fac_corr_mat = np.corrcoef(fac_disp, rowvar = True)
+        # extract relevant row from cov matrix
+        ref_lmk_index = [i for i, lmk in enumerate(landmarks) if config['ref_lmk']==lmk]
+        fac_corr = fac_corr_mat[ref_lmk_index][0]
+
+        fac_area = config['ref_area'] / (facew * faceh)
+
+        # if verbose:
+        #     logger.info("Face area: {}".format(fac_area))
+        #     logger.info("Face Displacement Median: {}".format(str(fac_disp_median)))
+        #     logger.info("Face Displacement Mean: {}".format(str(fac_disp_mean)))
+
+        fac_features1 =  np.multiply(fac_area * fac_disp_median, (1. - fac_corr))
+        fac_features2 =  np.multiply(fac_area * fac_disp_mean, (1. - fac_corr))
+
+#         base_fac_features = np.dot(fac_area * fac_disp_median, (1. - fac_corr))
+
+        fac_features_dict = {}
+        for i, landmark in enumerate(landmarks):
+            fac_features_dict['fac_features_mean_{}'.format(landmark)] = [fac_features2[i]]
+            raw_variable_map = 'fac_tremor_median_{}'.format(landmark)
+            fac_features_dict[r_config.raw_feature[raw_variable_map]] = [fac_features1[i]]
+
+            fac_features_dict['fac_disp_median_{}'.format(landmark)] = [fac_disp_median[i]]
+            fac_features_dict['fac_corr_{}'.format(landmark)] = [fac_corr[i]]
+
+        fac_features_dict[r_config.err_reason] = ['']
+        data = pd.DataFrame.from_dict(fac_features_dict)
+        logger.info('Concluded computing tremor features')
+
+        return data
+
+    except Exception as e:
+        logger.error('Error computing tremor features: {}'.format(str(e)))
+        return empty_frame(landmarks, r_config, str(e))
+
+def empty_frame(landmarks, r_config, error_reason):
+    fac_features_dict = {}
+    for i, landmark in enumerate(landmarks):
+        raw_variable_map = 'fac_tremor_median_{}'.format(landmark)
+        fac_features_dict[r_config.raw_feature[raw_variable_map]] = [np.nan]
+
+        fac_features_dict['fac_features_mean_{}'.format(landmark)] = [np.nan]
+        fac_features_dict['fac_disp_median_{}'.format(landmark)] = [np.nan]
+        fac_features_dict['fac_corr_{}'.format(landmark)] = [np.nan]
+
+    fac_features_dict[r_config.err_reason] = [error_reason]
+    empty_frame = pd.DataFrame.from_dict(fac_features_dict)
+    return empty_frame
+
+def fac_tremor_process(video_uri,out_dir,r_config, model_output=False):
+    """
+    processing input videos
+    """
+    try:
+        logger.info('filtering path: ',video_uri,out_dir)
+        input_loc, out_loc, fl_name = ut.filter_path(video_uri, out_dir)
+        of_csv_path = glob.glob(join(out_loc, fl_name + '_OF_video_features/*.csv'))
+
+        if len(of_csv_path)>0:
+            of_csv = of_csv_path[0]
+            df_of = pd.read_csv(of_csv, error_bad_lines=False)
+
+            logger.info('Processing Output file {} '.format(os.path.join(out_loc, fl_name)))
+
+            feats = compute_features(of_csv_path , df_of, r_config)
+            if model_output:
+                result = score(feats, r_config)
+                feats = pd.concat([feats, result], axis=1)
+
+                ut.output_audio_feature(feats, new_out_base_dir, '/'+fac_dir, fac_ext)
+
+    except Exception as e:
+        logger.error('Failed to process video file')