eDNSalModel.py

import numpy as np 
import scipy as sp 
from scipy import misc 
import time 
import logging

from liblinearutil import predict 

from features import eDN_features, dist_to_cntr_features, whiten_features

EDN_INSIZE = (512, 384) 

class EDNSaliencyModel(object): 
    
    def __init__(self, descriptions, svmModel, biasToCntr=False):  
        self.descriptions = descriptions 
        self.svm = svmModel['svm'] 
        self.whitenParams = svmModel['whitenParams'] 
        self.biasToCntr = biasToCntr 
        nFeatures = np.sum([d['desc'][-1][0][1]['initialize']['n_filters'] 
                        for d in self.descriptions if d != None])
        diff = self.svm.get_nr_feature() - nFeatures
        if diff != 0:
            if diff == 1 and not biasToCntr: 
                raise ValueError("The number of features in eDN and svm "
                                 "models does not match! Is the center bias "
                                 "correctly set?")

    def saliency(self, img, normalize=True): 
        """Computes eDN saliency map for single image or image sequence""" 
    
        descs = self.descriptions 

        # rescale image to typical input size 
        imgSize = img.shape[:2]  
        rescale_factor = 0.5*EDN_INSIZE[0]/max(imgSize) + \
                         0.5*EDN_INSIZE[1]/min(imgSize) 

        # single image: 
        # img = misc.imresize(image, rescale_factor, 'bicubic') 
        
        # image sequence (or single image) 
        scaledImg = np.zeros([a*rescale_factor for a in imgSize] + [img.shape[2]])
        for j in xrange(img.shape[2]/3):
            scaledImg[:,:,j*3:j*3+3] = misc.imresize(img[:,:,j*3:j*3+3],
                                                     rescale_factor, 'bicubic')
        img = scaledImg

        # compute eDN features for description(s) 
        t1 = time.time()
        fMapEDN, fMapSize = eDN_features(img, descs) 
        t2 = time.time()
        logging.info("Feature computation took %0.3fs" % (t2-t1))

        if self.biasToCntr:  
            fMapCntr = dist_to_cntr_features(img, fMapSize) 
            fMap = np.hstack((fMapCntr, fMapEDN)) 
        else: 
            fMap = fMapEDN 

        fMapW, fwParams = whiten_features(fMap, self.whitenParams) 

        # SVM prediction 
        t1 = time.time()
        bs, pAcc, pred = predict([], fMapW.tolist(), self.svm, options="-q") 
        t2 = time.time()
        logging.info("Prediction took %0.3fs" % (t2-t1))
        pred = np.array(pred) 

        # reshaping and upscaling
        pred = pred.reshape(fMapSize, order='F')
        predLarge = sp.ndimage.interpolation.zoom(pred, 
            (imgSize[0]/float(pred.shape[0]),
            imgSize[1]/float(pred.shape[1])))

        # normalization
        if normalize: 
            rescaled = (255.0 / (predLarge.max()-predLarge.min()) * 
                       (predLarge-predLarge.min())).astype(np.uint8) 
            return rescaled
        else:
            return predLarge