def generate_data_trip(img_data, C, backend,draw_flag = False): X, Y, img_data= data_generators.get_anchor_gt_trip(img_data, C, backend) P_rpn = model_rpn.predict_on_batch(X) R = roi_helpers.rpn_to_roi(P_rpn[0], P_rpn[1], C, K.image_dim_ordering(), use_regr=True, overlap_thresh=0.7, max_boxes=300) # note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format X2, Y1, Y2, Y_view = roi_helpers.calc_iou_new(R, img_data, C, class_mapping) pos_samples = np.where(Y_view[0, :, -1] != 20) if C.num_rois > 1: if len(pos_samples) < C.num_rois: selected_pos_samples = pos_samples[0].tolist() else: selected_pos_samples = np.random.choice(pos_samples[0], C.num_rois, replace=False).tolist() R,Y_view = roi_helpers.prep_flip(X2[:,selected_pos_samples,:],Y_view[:,selected_pos_samples,:],C) if draw_flag: Im = cv2.imread(img_data['filepath']) key = img_data['bboxes'][0]['class'] azimuth = img_data['bboxes'][0]['azimuth'] bbox = np.array([[img_data['bboxes'][0]['x1'],img_data['bboxes'][0]['y1'],img_data['bboxes'][0]['x2'],img_data['bboxes'][0]['y2']]]) img = img_helpers.draw_bbox(Im, bbox, 0, [azimuth], 1, class_mapping_inv, key) img_helpers.display_image(img,0) return X,R,Y_view
global im_id im_id+=1 return im_id siam_flip = True if siam_flip and iter_num % C.siam_iter_frequancy == 0 : idx_orig = pos_samples.tolist() # idx_flip = pos_samples_flip.tolist() xSize = P_rpn[0].shape[1] R_tmp = X2[:,idx_orig,:] Y_tmp = Y_view[:,idx_orig,:] ## make sure that none of the ROI is neg idx_pos = xSize - R_tmp[0,:,2] + R_tmp[0,:,0] > 0 R_orig,Y_orig = roi_helpers.prep_flip(R_tmp[:,idx_pos,:],Y_tmp[:,idx_pos,:],C) R_flip = copy.deepcopy(R_orig) R_flip[0,:,0] += xSize - R_flip[0,:,2] if np.unique(Y_orig[0,:,-1]).shape[0] > 1 or sum(idx_pos) < 5: continue model_classifier.save_weights(weight_path_tmp) model_view_only.load_weights(weight_path_tmp, by_name=True) loss_before = model_flip.predict([X, R_orig,X_flip,R_flip,Y_orig]) # model_flip.predict([X, R_orig,X,R_orig]) _, _, out_az = model_classifier.predict([X, R_orig]) _, _, out_az_flip = model_classifier.predict([X_flip, R_flip]) gt_label = np.argmax(Y1[:, idx_orig, :], axis=2) az=[] az_flip=[]