Rebecca Mason
http://cs.brown.edu/people/rebecca

Resources for "Nonparametric Method for Data-driven Image Captioning":
 
We use the same 1000 query images as used by:
http://www.cs.sunysb.edu/~pkuznetsova/generation/cogn/captions.html

However, we noticed that some of the captions had train and test contamination, so we removed them. Those captions correspond with the following images in the above list (note these positions are 1-indexed because Matlab). 

4	28	35	86	87	133	137	161	196	211	250	356	366	399	403	413	455	456	457	458	477	514	557	586	609	688	689	759	769	791	839	841	866	869	900	903	917	967	968

Captions are provided in .sgm format for BLEU evaluation. We include the following:

human (reference) captions:
collective_ref.sgm
collective_gold.sgm (these are the same captions, but you need both files to run BLEU)

our system captions:
kl = kl divergence extraction
sb = SumBasic extraction

kl_k25.sgm is the system that we used for the human evaluations.
sb_k25.sgm is the SumBasic extraction version of the same system.
sb_k500.sgm is the highest performing system according to BLEU evaluation.

comparison captions:
scene_attributes.sgm is the Scene Attributes captions from Patterson et al. 2014, and the k=1 baseline for our system
collective.sgm is the Collective system from Kuznetsova et al. 2012 with cognitive phrases, retrieved from the website above. 


