COCO JSON dataset masks parser

PythonAPI/extra/README.md

@@ -0,0 +1,43 @@
+
+
+
+# MASKS_PARSER.PY
+
+
+
+## Notes:
+
+- This program performs a .JSON segmentation masks parsing and applies conversion formats: RLE <-> Polygon.
+- Note #1: Conversion only applied to COCO-format annotations "iscrowd==1"
+- Note #2: Script only tested over COCO dataset formats
+- Note #3: This script overwrites input dataset with same name as input (to avoid redundancy)
+- Note #4: Attached sample dataset is a sub-sample of COCO Dataset 2017 for 1 single image with one (only) multi-object annotation (is_crowd= 1). See:  http://images.cocodataset.org/zips/val2017.zip
+- Author: Aleix Figueres ([email protected])
+
+
+
+## Instructions:
+
+1. Go to "extra" (source):
+
+  	`cd ./extra
+2. Modify .JSON dataset input path on "main" function:
+
+  	 ` # Input dataset path
+  	 DATASET_PATH = './samples/instances_val2017_1img-sample.json'`
+
+3. Comment/uncomment desired conversion format:
+
+  	` # SAMPLE #1: Convert RLE to Polygon
+  	RLE2poly(DATASET_PATH) `<br>
+  	` # SAMPLE #2: Convert Polygon to RLE
+  	poly2RLE(DATASET_PATH) `
+
+- Note 5: when running both (RLE2poly + poly2RLe) output will not be identical (bit-a-bit) to input dataset due to encoding data loosing but after comparing output vs input masks, quite-identical shapes got obtained
+
+3. Run:
+
+    `python masks_parser.py`
+
+4. Check outputs on "DATASET_PATH"
+

PythonAPI/extra/masks_parser.py

@@ -0,0 +1,139 @@
+
+"""masks_parser.py
+This program performs a .JSON segmentation masks parsing and applies conversion formats: RLE <-> Polygon.
+Note: Conversion only applied to COCO-format annotations "iscrowd==1"
+"""
+__author__ = "[email protected]"
+
+import encodings
+import numpy as np
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+from PIL import Image
+import requests
+import cv2
+import json
+from itertools import groupby
+import sys
+from pathlib import Path
+import os
+# Get current directory
+
+cur_dir = os.getcwd()
+sys.path.append(os.path.join(cur_dir, "../pycocotools"))  # To find local version
+
+from pycocotools.coco import COCO
+import pycocotools._mask as _mask
+
+## Global defs       
+def RLE2poly(file_path):
+    """
+    Converts binary mask format to RLE and applies this conversion on same input file
+    :param file_path: COCO annotation path directory (dataset)
+    :return: Null (returns overwritten file)
+    """
+    try:
+        coco_annotation = COCO(annotation_file=file_path)
+
+        for x in range(0,len(coco_annotation.dataset['annotations']),1):
+            # Conversion only applicable for COCO's "iscrowd" annotations
+            if(int(coco_annotation.dataset['annotations'][x]['iscrowd']) == 1):
+                RLE_mask = coco_annotation.annToMask(coco_annotation.dataset['annotations'][x])
+                list_test = []
+                polygon_decoded= polygonFromMask(RLE_mask)
+                list_test.append(polygon_decoded)
+                # Rewrite dataset
+                with open(file_path, "r",encoding='utf-8') as jsonFile:
+                    data = json.load(jsonFile)
+                    data['annotations'][x]['segmentation'] = list_test
+
+                with open(file_path, "w",encoding='utf-8') as jsonFile:
+                    json.dump(data, jsonFile)
+
+        print("Conversion DONE -> RLE2poly")
+    except:
+        print("Error: Incorrect RLE2poly conversion. System aborted")
+        sys.exit(0)
+
+def poly2RLE(file_path):
+    """
+    Converts polygon object format to RLE and applies this conversion on same input file
+    :param file_path: COCO annotation path directory (dataset)
+    :return: Null (returns overwritten file)
+    """
+    try:
+        coco_annotation = COCO(annotation_file=file_path)
+
+        for x in range(0,len(coco_annotation.dataset['annotations']),1):
+            # Conversion only applicable for COCO's "iscrowd" annotations
+            if(int(coco_annotation.dataset['annotations'][x]['iscrowd']) == 1):
+                RLE_candidate = coco_annotation.annToMask(coco_annotation.dataset['annotations'][x])
+                list_test = []
+                RLE_encoded= binary_mask_to_rle(RLE_candidate)
+                list_test.append(RLE_encoded)
+                #annotation = coco_annotation.dataset['annotations'][x]['segmentation'][0] = RLE_encoded
+
+                # Rewrite dataset
+                with open(file_path, "r",encoding='utf-8') as jsonFile:
+                    data = json.load(jsonFile)
+                    data['annotations'][x]['segmentation'] = RLE_encoded
+
+                with open(file_path, "w",encoding='utf-8') as jsonFile:
+                    json.dump(data, jsonFile)
+
+        print("Conversion DONE -> poly2RLE")
+    except:
+        print("Error: Incorrect poly2RLE conversion. System aborted")
+        sys.exit(0)
+
+def polygonFromMask(maskedArr): 
+    """
+    Converts binary mask format to RLE
+    Source: https://github.com/hazirbas/coco-json-converter/blob/master/generate_coco_json.py
+    :param maskedArr: binary mask
+    :return: polygon mask  ([x, y, w, h], area )
+    """
+    contours, _ = cv2.findContours(maskedArr, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
+
+    segmentation = []
+    for contour in contours:
+        # Valid polygons have >= 6 coordinates (3 points)
+        if contour.size >= 6:
+            segmentation.append(contour.flatten().tolist())
+    RLEs = _mask.frPyObjects(segmentation, maskedArr.shape[0], maskedArr.shape[1])
+    RLE = _mask.merge(RLEs)
+    # RLE = mask.encode(np.asfortranarray(maskedArr))
+    area = _mask.area(RLEs)
+    [x, y, w, h] = cv2.boundingRect(maskedArr)
+
+    return segmentation[0] 
+
+
+def binary_mask_to_rle(binary_mask):
+    """
+    Converts binary mask format to RLE
+    Source: https://stackoverflow.com/a/49547872
+    :param binary_mask: binary mask
+    :return: RLE mask 
+    """
+    rle = {'counts': [], 'size': list(binary_mask.shape)}
+    counts = rle.get('counts')
+    for i, (value, elements) in enumerate(groupby(binary_mask.ravel(order='F'))):
+        if i == 0 and value == 1:
+            counts.append(0)
+        counts.append(len(list(elements)))
+    return rle
+
+
+
+
+if __name__ == "__main__":
+    # Input dataset path
+    DATASET_PATH = './samples/instances_val2017_1img-sample.json'
+
+    # SAMPLE #1: Convert RLE to Polygon
+    RLE2poly(DATASET_PATH)
+
+    # SAMPLE #2: Convert Polygon to RLE
+    poly2RLE(DATASET_PATH)