sherlock project

 

Github link

https://github.com/mhelhoseiny/sherlock


Download Benchmarked Developed in [1] 


1) 6DS benchmark https://www.dropbox.com/s/lwbg1klshqlc52k/6DS_dataset.zip (2.4 GB)
(28,000 images, 186 unique facts) 
wit the training and testing splits
Fact Recognition Top 1 Accuracy (our method): 
69.63% 
Fact Recognition MAP/MAP100 (our method): 34.86%/ 50.68%
2) 
 LSC (Large Scale benchmark) 
  (814K images, 202K unique  facts) 
with the training and testing splits

part 1,LSC_dataset.tar.gz.aa  https://www.dropbox.com/s/o9nr3l7h1pbx7en/LSC_dataset.tar.gz.aa?dl=0 (11.72 GB):
part 2, LSC_dataset.tar.gz.ab https://www.dropbox.com/s/ep6oopfpykmtuql/LSC_dataset.tar.gz.ab?dl=0 (8.73 GB):
After download
cat LSC_dataset.tar.gz.* > LSC_dataset.tar.gz

Then extract 
LSC_dataset.tar.gz

Models
1) Model 2 trained on LSC benchmark  caffemodel  deploy_prototxt


Feel free to contact for any questions. 







References
 [1] Mohamed Elhoseiny, Scott Cohen, Walter Chang, Brian Price, Ahmed Elgammal, 
Sherlock: Scalable Fact Learning in Images, AAAI, 2017, acceptance rate (24%).
 

[2] Mohamed Elhoseiny, Scott Cohen, Walter Chang, Brian Price, Ahmed Elgammal, 
Automatic Annotation of Structured Facts in Images, ACL Proceedings of the  Vision&Language Workshop, 2016 (long paper)