Some changes to allow training with kaggle data

normalization/blazepose_mapping.py

@@ -61,20 +61,25 @@ def map_blazepose_keypoint(column):
     return f"{mapped}_{hand}{suffix}"
 
 
-def map_blazepose_df(df):
+def map_blazepose_df(df, rename=True):
+    to_drop = []
+    if rename:
+        renamings = {}
+        for column in df.columns:
+            mapped_column = map_blazepose_keypoint(column)
+            if mapped_column:
+                renamings[column] = mapped_column
+            else:
+                to_drop.append(column)
+        df = df.rename(columns=renamings)
+
     for index, row in df.iterrows():
 
+        sequence_size = len(row["leftEar_Y"])
         lsx = row["leftShoulder_X"]
         rsx = row["rightShoulder_X"]
         lsy = row["leftShoulder_Y"]
         rsy = row["rightShoulder_Y"]
-        # convert all to list
-        lsx = lsx[1:-1].split(",")
-        rsx = rsx[1:-1].split(",")
-        lsy = lsy[1:-1].split(",")
-        rsy = rsy[1:-1].split(",")
-        sequence_size = len(lsx)
-
         neck_x = []
         neck_y = []
         # Treat each element of the sequence (analyzed frame) individually
@@ -84,4 +89,5 @@ def map_blazepose_df(df):
         df.loc[index, "neck_X"] = str(neck_x)
         df.loc[index, "neck_Y"] = str(neck_y)
 
-    return df
+    df.drop(columns=to_drop, inplace=True)
+    return df

normalization/main.py

@@ -5,30 +5,30 @@ import pandas as pd
 from normalization.hand_normalization import normalize_hands_full
 from normalization.body_normalization import normalize_body_full
 
-DATASET_PATH = './data/wlasl'
+DATASET_PATH = './data/processed'
 # Load the dataset
-df = pd.read_csv(os.path.join(DATASET_PATH, "WLASL100_train.csv"), encoding="utf-8")
+df = pd.read_csv(os.path.join(DATASET_PATH, "spoter_train.csv"), encoding="utf-8")
 
 print(df.head())
 print(df.columns)
 
 # Retrieve metadata
-video_size_heights = df["video_height"].to_list()
-video_size_widths = df["video_width"].to_list()
+# video_size_heights = df["video_height"].to_list()
+# video_size_widths = df["video_width"].to_list()
 
 # Delete redundant (non-related) properties
-del df["video_height"]
-del df["video_width"]
+# del df["video_height"]
+# del df["video_width"]
 
 # Temporarily remove other relevant metadata
 labels = df["labels"].to_list()
-video_fps = df["fps"].to_list()
+signs = df["sign"].to_list()
+
 del df["labels"]
-del df["fps"]
-del df["split"]
-del df["video_id"]
-del df["label_name"]
-del df["length"]
+del df["sign"]
+del df["path"]
+del df["participant_id"]
+del df["sequence_id"]
 
 # Convert the strings into lists
 
@@ -41,7 +41,7 @@ for column in df.columns:
 
 # Perform the normalizations
 df = normalize_hands_full(df)
-df, invalid_row_indexes = normalize_body_full(df)
+# df, invalid_row_indexes = normalize_body_full(df)
 
 # Clear lists of items from deleted rows
 # labels = [t for i, t in enumerate(labels) if i not in invalid_row_indexes]
@@ -49,6 +49,6 @@ df, invalid_row_indexes = normalize_body_full(df)
 
 # Return the metadata back to the dataset
 df["labels"] = labels
-df["fps"] = video_fps
+df["sign"] = signs
 
-df.to_csv(os.path.join(DATASET_PATH, "wlasl_train_norm.csv"), encoding="utf-8", index=False)
+df.to_csv(os.path.join(DATASET_PATH, "spoter_train_norm.csv"), encoding="utf-8", index=False)