swmaestro_image_retrieval_evaluator

software maestro 2018 backend - image retrieval contest evaluation server

• score leader board : http://eval.buzzni.net:31000/leader_board
• data set : https://www.dropbox.com/s/pbui2jycsuimomu/fashion_retrieval_img.tar.gz?dl=0
  • query folder : query image
  • compare folder : search target image
• how to make evaluation?
  1. iterate all query image and calculate distance with all compare image.
  2. sort compare image by distance (asc)
  3. upload the result (160 query)
• evaluation server : http://eval.buzzni.net:31000
• api description : http://eval.buzzni.net:31000/apidocs/#/
• sample upload evaluation code

system_result_dict = {}
for each in glob("/home/maestro_2018/test/img/query/*"):
  fname = each.split("/")[-1]
  score_dict = {}
  for other in glob("/home/maestro_2018/test/img/compare/*"):
      fname2 = other.split("/")[-1]
      dist = cosine(img_feature_dict[fname], img_feature_dict[fname2])
      score_dict[fname2] = dist
  sorted_list = sorted(score_dict.items(), key=lambda (k,v):(v,k), reverse=False)
  qid = fname.split("_")[-1].split(".")[0]
  system_result_dict[qid] = map(lambda i : i[0], sorted_list[:20])
        
name = 'buzzni-200'
nickname = 'buzzni-200'
r = requests.post('http://eval.buzzni.net:31000', json={"pred_result": system_result_dict,'name':name, 'nickname':nickname})
print r.json()

• 버즈니에서 개발자 채용을 하고 있습니다 관심 있는 분들은 recruit@buzzni.com 로 연락 주세요. 전문연구요원, 보충역 가능합니다.