utils_ade20k.py

from PIL import Image
import matplotlib._color_data as mcd
import cv2
import ipdb
import json
import numpy as np
import os

_NUMERALS = '0123456789abcdefABCDEF'
_HEXDEC = {v: int(v, 16) for v in (x+y for x in _NUMERALS for y in _NUMERALS)}
LOWERCASE, UPPERCASE = 'x', 'X'
def rgb(triplet):
    return _HEXDEC[triplet[0:2]], _HEXDEC[triplet[2:4]], _HEXDEC[triplet[4:6]]

def loadAde20K(file):
    fileseg = file.replace('.jpg', '_seg.png');
    with Image.open(fileseg) as io:
        seg = np.array(io);

    # Obtain the segmentation mask, bult from the RGB channels of the _seg file
    R = seg[:,:,0];
    G = seg[:,:,1];
    B = seg[:,:,2];
    ObjectClassMasks = (R/10).astype(np.int32)*256+(G.astype(np.int32));


    # Obtain the instance mask from the blue channel of the _seg file
    Minstances_hat = np.unique(B, return_inverse=True)[1]
    Minstances_hat = np.reshape(Minstances_hat, B.shape)
    ObjectInstanceMasks = Minstances_hat


    level = 0
    PartsClassMasks = [];
    PartsInstanceMasks = [];
    while True:
        level = level+1;
        file_parts = file.replace('.jpg', '_parts_{}.png'.format(level));
        if os.path.isfile(file_parts):
            with Image.open(file_parts) as io:
                partsseg = np.array(io);
            R = partsseg[:,:,0];
            G = partsseg[:,:,1];
            B = partsseg[:,:,2];
            PartsClassMasks.append((np.int32(R)/10)*256+np.int32(G));
            PartsInstanceMasks = PartsClassMasks
            # TODO:  correct partinstancemasks

            
        else:
            break

    objects = {}
    parts = {}

    attr_file_name = file.replace('.jpg', '.json')
    if os.path.isfile(attr_file_name):
        with open(attr_file_name, 'r') as f:
            input_info = json.load(f)

        contents = input_info['annotation']['object']
        instance = np.array([int(x['id']) for x in contents])
        names = [x['raw_name'] for x in contents]
        corrected_raw_name =  [x['name'] for x in contents]
        partlevel = np.array([int(x['parts']['part_level']) for x in contents])
        ispart = np.array([p>0 for p in partlevel])
        iscrop = np.array([int(x['crop']) for x in contents])
        listattributes = [x['attributes'] for x in contents]
        polygon = [x['polygon'] for x in contents]
        for p in polygon:
            p['x'] = np.array(p['x'])
            p['y'] = np.array(p['y'])

        objects['instancendx'] = instance[ispart == 0]
        objects['class'] = [names[x] for x in list(np.where(ispart == 0)[0])]
        objects['corrected_raw_name'] = [corrected_raw_name[x] for x in list(np.where(ispart == 0)[0])]
        objects['iscrop'] = iscrop[ispart == 0]
        objects['listattributes'] = [listattributes[x] for x in list(np.where(ispart == 0)[0])]
        objects['polygon'] = [polygon[x] for x in list(np.where(ispart == 0)[0])]


        parts['instancendx'] = instance[ispart == 1]
        parts['class'] = [names[x] for x in list(np.where(ispart == 1)[0])]
        parts['corrected_raw_name'] = [corrected_raw_name[x] for x in list(np.where(ispart == 1)[0])]
        parts['iscrop'] = iscrop[ispart == 1]
        parts['listattributes'] = [listattributes[x] for x in list(np.where(ispart == 1)[0])]
        parts['polygon'] = [polygon[x] for x in list(np.where(ispart == 1)[0])]

    return {'img_name': file, 'segm_name': fileseg,
            'class_mask': ObjectClassMasks, 'instance_mask': ObjectInstanceMasks, 
            'partclass_mask': PartsClassMasks, 'part_instance_mask': PartsInstanceMasks, 
            'objects': objects, 'parts': parts}

def plot_polygon(img_name, info, show_obj=True, show_parts=False):
    colors = mcd.CSS4_COLORS
    color_keys = list(colors.keys())
    all_objects = []
    all_poly = []
    if show_obj:
        all_objects += info['objects']['class']
        all_poly += info['objects']['polygon']
    if show_parts:
        all_objects += info['parts']['class']
        all_poly += info['parts']['polygon']

    img = cv2.imread(img_name)
    thickness = 5
    for it, (obj, poly) in enumerate(zip(all_objects, all_poly)):
        curr_color = colors[color_keys[it % len(color_keys)] ]
        pts = np.concatenate([poly['x'][:, None], poly['y'][:, None]], 1)[None, :]
        color = rgb(curr_color[1:])
        img = cv2.polylines(img, pts, True, color, thickness)
    return img