| BOORU CHARS dataset can be used for: | |
| - NN training for tag detection (as Deep Danbooru and better) and/or object detection | |
| - scene scale and composition classification based on key objects (e.g. face) detected | |
| I used notAI-tech NudeNet detector (github.com/notAI-tech/NudeNet) | |
| # rewrote logging and painting of detected objects | |
| def censor(self, img_path, out_path ): | |
| image = cv2.imread(img_path) | |
| ih, iw, _ = image.shape | |
| boxes = self.detect(img_path) | |
| i = 0 | |
| for box in boxes: | |
| if 'FACE' in box['label'] : colr=(0, 255, 0) | |
| if 'EXPOSED' in box['label'] : colr=(0, 0, 255) | |
| if 'COVERED' in box['label'] : colr=(255, 0, 0) | |
| lnw = int(min(ih, iw)/48*(box['score']-0.3)) | |
| print(img_path+';'+box['label']+';'+str(round(box['score'],2))+';'+str(box['box'][0])+';'\ | |
| +str(box['box'][1])+';'+str(box['box'][2]-box['box'][0])+';'+str(box['box'][3]-box['box'][1]),flush = True) | |
| if 'FACE' in box['label'] : | |
| x = box['box'][0] | |
| y = box['box'][1] | |
| w = box['box'][2] - box['box'][0] | |
| h = box['box'][3] - box['box'][1] | |
| crop_img = image[max(0,int(y-0.6*h)): int(y+1.1*h), max(0,int(x-0.35*w)): min(iw,int(x+1.35*w))] | |
| i = 1 | |
| image = cv2.rectangle( | |
| image, (box['box'][0], box['box'][1]), (box['box'][2], box['box'][3]), colr, max(lnw, 4) | |
| ) | |
| cv2.putText(image, box['label'] + ' ' + str(round(box['score'],2)), | |
| (box['box'][0], box['box'][1]-int(ih*0.005)), cv2.FONT_HERSHEY_SIMPLEX, | |
| ih*0.0004, | |
| (0,0,0), 2, 2) | |
| cv2.imwrite(out_path, image) | |
| # main loop | |
| if __name__ == "__main__": | |
| m = Detector() | |
| for fname in os.listdir(sys.argv[1]): | |
| m.censor(sys.argv[1]+fname,sys.argv[2]+fname ) | |
| With volumes 2020-3x4 and 2020-1x2 processed I got 200.887 detections listed in BCN_detect.tsv | |
| FNAME; class OBJ; probability PROB coordinates X;Y;W;H | |
| and also copies of sample pictures with objects shown. | |
| Then I used DB (PL/SQL stored proc) for | |
| - Non Maximum Suppression | |
| - assembling algorithm (in polar coordinates), started from FACE which | |
| * find ARMPITS & BREASTS (with suitable size and position) | |
| * then BELLY (using all findings above) | |
| * then GENITALIA & ANUS (again using findings available) | |
| * and then FEETS (most trucky because of distance and possible vector change) | |
| All findings associated with faces listed in BCN_lineup.tsv (78.184 rows) where: | |
| 'FNAME' - file name | |
| 'FACE_ID' - face hash inside file | |
| 'PROB','X','Y','W','H' - face parameters | |
| 'OBJ' - assembled object type simplified ('BRST','ARMP','BELL' etc) | |
| 'OPROB','OX','OY','OW','OH' - object parameters | |
| Using listing I can recolor related object and join it center-to-center using code | |
| # PYTHON | |
| import cv2 | |
| import pandas as pd | |
| prev_fname = 'NONE' | |
| prev_oname = 'NONE' | |
| prev_hashid = 0 | |
| data = pd.read_csv('in.lst',sep=';', decimal=',',index_col='IDX') | |
| #"IDX";"FNAME";"HASHID";"PROB";"X";"Y";"W";"H";"OBJ";"OPROB";"OX";"OY";"OW";"OH";"ONAME" | |
| for i, row in data.iterrows(): | |
| print(str(i)+' '+row['FNAME']) | |
| if row['FNAME']!=prev_fname: | |
| image = cv2.imread(row['FNAME']) | |
| ih, iw, _ = image.shape | |
| print(str(i) + ' RD ' + row['FNAME']) | |
| if row['ONAME']!=prev_oname and prev_oname!='NONE': | |
| cv2.imwrite(prev_oname, prev_image) | |
| print(str(i)+' WR '+prev_oname) | |
| prev_oname=row['ONAME'] | |
| prev_fname=row['FNAME'] | |
| prev_image = image | |
| lnw = int(min(ih, iw) / 48 * (row['PROB'] - 0.3)) | |
| prev_image = cv2.rectangle( prev_image, (row['X'], row['Y']),\ | |
| (row['X']+row['W'], row['Y']+row['H']), (255,255,255), max(lnw, 4) ) | |
| lnwr = int(min(ih, iw) / 48 * (row['OPROB'] - 0.3)) | |
| if row['OBJ']=='BRST': | |
| colr = (255, 153, 0) # light blue | |
| lnw = int(lnwr*0.66) | |
| if row['OBJ'] in ('BELL','XXXX'): | |
| colr = (0, 153, 255) # orange | |
| lnw = int(lnwr*0.66) | |
| if row['OBJ']=='ARMP': | |
| colr = (153,0,153) # violet | |
| lnw = int(lnwr*0.33) | |
| if row['OBJ']=='FEET': | |
| colr = (51, 153, 102) # green | |
| lnw = int(lnwr*0.33) | |
| prev_image = cv2.line( prev_image, (int(row['X']+row['W']/2), int(row['Y']+row['H']/2)),\ | |
| (int(row['OX']+row['OW']/2), int(row['OY']+row['OH']/2)),colr, max(lnw, 4) ) | |
| prev_image = cv2.rectangle( prev_image, (row['OX'], row['OY']),\ | |
| (row['OX']+row['OW'], row['OY']+row['OH']), colr, max(lnwr, 4) ) | |
| cv2.imwrite(prev_oname, prev_image) | |
| print(str(i) + ' WR ' + prev_oname) | |
| You can find ~4000 selected examples in archives 2020-1x2_O.zip and 2020-3x4_O.zip with: | |
| - 5 or more relations (good complex detection scenarios) | |
| - some EXPOSED objects with high probabilities (possible fails in manual cleanup) | |
| Much more examples omitted because on release size. | |
| Much more efforts has to be done for NudeNet (better detection) and for me (better assembling). | |
| Initial task "scene scale and composition classification" will require | |
| - much more (100.000+ ?) good complex detections | |
| - data mining (attribute importance, clustering) with Oracle DBMS_DATA_MINING or python equivalent | |