''' ------------------------------------------------------------------------------------------------------ ,@@@@@@ @@@@@@@@@@@ @@@ @@@@@@@@@@@@ @@@@@@@@@@@ @@@@@@@@@@@@@ @@@@@@@@@@@@@@ @@@@@@@/ ,@@@@@@@@@@@@@ /@@@@@@@@@@@@@@@ @@@@@@@@ @@@@@@@@@@@@@@@@@@@@@@@@ @@@@@ @@@@@@@@ @@@@@ ,@@@ @@@@& @@@@@@. @@@@ @@@ @@@@@@@@@/ @@@@@ ,@@@. @@@@@@((@ @@@@( //@@@ ,, @@@@ @@@@@ @@@( @@@@@@@ @@@ @ @@@@@@@@# @@@@@@@@@@@@@@@@@ @@@@@@@@@@@@@( RANSAC 3D By: Summer#2406 (Main Algorithm Engineer), Pupil Labs (pye3d), PallasNeko (Optimization) Algorithm App Implimentations By: Prohurtz#0001, qdot (Inital App Creator) Copyright (c) 2022 EyeTrackVR <3 ------------------------------------------------------------------------------------------------------ ''' import cv2 import numpy as np def ellipse_model(data, y, f): """ There is no need to make this process a function, since making the process a function will slow it down a little by calling it. The results may be slightly different from the lambda version due to calculation errors derived from float types, but the calculation results are virtually the same. a = 1.0,b = P[0],c = P[1],d = P[2],e = P[3],f = P[4] :param data: :param y: np.c_[d, e, a, c, b] :param f: f == P[4, 0] :return: this_return == np.array([ellipse_model(x, y) for (x, y) in data ]) """ return data.dot(y) + f # @profile def fit_rotated_ellipse_ransac(data: np.ndarray, rng: np.random.Generator, iter=100, sample_num=10, offset=80 # 80.0, 10, 80 ): # before changing these values, please read up on the ransac algorithm # However if you want to change any value just know that higher iterations will make processing frames slower effective_sample = None # The array contents do not change during the loop, so only one call is needed. # They say len is faster than shape. # Reference url: https://stackoverflow.com/questions/35547853/what-is-faster-python3s-len-or-numpys-shape len_data = len(data) if len_data < sample_num: return None # Type of calculation result ret_dtype = np.float64 # Sorts a random number array of size (iter,len_data). After sorting, returns the index of sample_num random numbers before sorting. # If the array size is less than about 100, this is faster than rng.choice. rng_sample = rng.random((iter, len_data)).argsort()[:, :sample_num] # or # I don't see any advantage to doing this. # rng_sample = np.asarray(rng.random((iter, len_data)).argsort()[:, :sample_num], dtype=np.int32) # I don't think it looks beautiful. # x,y,x**2,y**2,x*y,1,-1*x**2 datamod = np.concatenate( [data, data ** 2, (data[:, 0] * data[:, 1])[:, np.newaxis], np.ones((len_data, 1), dtype=ret_dtype), (-1 * data[:, 0] ** 2)[:, np.newaxis]], axis=1, dtype=ret_dtype) datamod_slim = np.array(datamod[:, :5], dtype=ret_dtype) datamod_rng = datamod[rng_sample] datamod_rng6 = datamod_rng[:, :, 6] datamod_rng_swap = datamod_rng[:, :, [4, 3, 0, 1, 5]] datamod_rng_swap_trans = datamod_rng_swap.transpose((0, 2, 1)) # These two lines are one of the bottlenecks datamod_rng_5x5 = np.matmul(datamod_rng_swap_trans, datamod_rng_swap) datamod_rng_p5smp = np.matmul(np.linalg.inv(datamod_rng_5x5), datamod_rng_swap_trans) datamod_rng_p = np.matmul(datamod_rng_p5smp, datamod_rng6[:, :, np.newaxis]).reshape((-1, 5)) # I don't think it looks beautiful. ellipse_y_arr = np.asarray( [datamod_rng_p[:, 2], datamod_rng_p[:, 3], np.ones(len(datamod_rng_p)), datamod_rng_p[:, 1], datamod_rng_p[:, 0]], dtype=ret_dtype) ellipse_data_arr = ellipse_model(datamod_slim, ellipse_y_arr, np.asarray(datamod_rng_p[:, 4])).transpose((1, 0)) ellipse_data_abs = np.abs(ellipse_data_arr) ellipse_data_index = np.argmax(np.sum(ellipse_data_abs < offset, axis=1), axis=0) effective_data_arr = ellipse_data_arr[ellipse_data_index] effective_sample_p_arr = datamod_rng_p[ellipse_data_index] return fit_rotated_ellipse(effective_data_arr, effective_sample_p_arr) # @profile def fit_rotated_ellipse(data, P): a = 1.0 b = P[0] c = P[1] d = P[2] e = P[3] f = P[4] # The cost of trigonometric functions is high. theta = 0.5 * np.arctan(b / (a - c), dtype=np.float64) theta_sin = np.sin(theta, dtype=np.float64) theta_cos = np.cos(theta, dtype=np.float64) tc2 = theta_cos ** 2 ts2 = theta_sin ** 2 b_tcs = b * theta_cos * theta_sin # Do the calculation only once cxy = b ** 2 - 4 * a * c cx = (2 * c * d - b * e) / cxy cy = (2 * a * e - b * d) / cxy # I just want to clear things up around here. cu = a * cx ** 2 + b * cx * cy + c * cy ** 2 - f cu_r = np.array([(a * tc2 + b_tcs + c * ts2), (a * ts2 - b_tcs + c * tc2)]) wh = np.sqrt(cu / cu_r) w, h = wh[0], wh[1] error_sum = np.sum(data) # print("fitting error = %.3f" % (error_sum)) return (cx, cy, w, h, theta) def RANSAC3D(self): f = False kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3)) thresh_add = 10 rng = np.random.default_rng() # Convert the image to grayscale, and set up thresholding. Thresholds here are basically a # low-pass filter that will set any pixel < the threshold value to 0. Thresholding is user # configurable in this utility as we're dealing with variable lighting amounts/placement, as # well as camera positioning and lensing. Therefore everyone's cutoff may be different. # # The goal of thresholding settings is to make sure we can ONLY see the pupil. This is why we # crop the image earlier; it gives us less possible dark area to get confused about in the # next step. if self.config.gui_circular_crop == True: if self.cct == 0: try: ht, wd = self.current_image_gray.shape[:2] radius = int(float(self.lkg_projected_sphere["axes"][0])) self.xc = int(float(self.lkg_projected_sphere["center"][0])) self.yc = int(float(self.lkg_projected_sphere["center"][1])) # draw filled circle in white on black background as mask mask = np.zeros((ht, wd), dtype=np.uint8) mask = cv2.circle(mask, (self.xc, self.yc), radius, 255, -1) # create white colored background color = np.full_like(self.current_image_gray, (255)) # apply mask to image masked_img = cv2.bitwise_and(self.current_image_gray, self.current_image_gray, mask=mask) # apply inverse mask to colored image masked_color = cv2.bitwise_and(color, color, mask=255 - mask) # combine the two masked images self.current_image_gray = cv2.add(masked_img, masked_color) except: pass else: self.cct = self.cct - 1 else: self.cct = 300 # Crop first to reduce the amount of data to process. newFrame2 = self.current_image_gray.copy() frame = self.current_image_gray # For measuring processing time of image processing # Crop first to reduce the amount of data to process. frame = frame[0:len(frame) - 5, :] # To reduce the processing data, first convert to 1-channel and then blur. # The processing results were the same when I swapped the order of blurring and 1-channelization. frame_gray = cv2.GaussianBlur(frame, (5, 5), 0) # this will need to be adjusted everytime hardware is changed (brightness of IR, Camera postion, etc)m min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(frame_gray) maxloc0_hf, maxloc1_hf = int(0.5 * max_loc[0]), int(0.5 * max_loc[1]) # crop 15% sqare around min_loc # frame_gray = frame_gray[max_loc[1] - maxloc1_hf:max_loc[1] + maxloc1_hf, # max_loc[0] - maxloc0_hf:max_loc[0] + maxloc0_hf] threshold_value = min_val + thresh_add _, thresh = cv2.threshold(frame_gray, threshold_value, 255, cv2.THRESH_BINARY) try: opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel) closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel) th_frame = 255 - closing except: # I want to eliminate try here because try tends to be slow in execution. th_frame = 255 - frame_gray contours, _ = cv2.findContours(th_frame, cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE) hull = [] # This way is faster than contours[i] # But maybe this one is faster. hull = [cv2.convexHull(cnt, False) for cnt in contours] for cnt in contours: hull.append(cv2.convexHull(cnt, False)) if not hull: # If empty, go to next loop pass try: cnt = sorted(hull, key=cv2.contourArea) maxcnt = cnt[-1] # ellipse = cv2.fitEllipse(maxcnt) ransac_data = fit_rotated_ellipse_ransac(maxcnt.reshape(-1, 2), rng) if ransac_data is None: # ransac_data is None==maxcnt.shape[0]