Image Question Answering Full Results

Reference: Mengye Ren, Ryan Kiros, Richard Zemel, "Exploring Models and Data for Image Question Answering", NIPS 2015.

DAQUAR-37

Dataset

References:

QAs: Mateusz Malinowski, Mario Fritz, "Towards a Visual Turing Challenge", NIPS 2014 Workshop on Learning Semantics. [ArXiv][link]

Images: Nathan Silberman, Pushmeet Kohli, Derek Hoiem, Rob Fergus, "Indoor Segmentation and Support Inference from RGBD Images", ECCV 2012. [link]

Notes:

1. Here we are only using DAQUAR-37 with one-word answers, a subset of the 37 object classes dataset.
2. Only test set results are rendered in the links below.

Individual Models

• 2-VIS+BLSTM
• VIS+LSTM
• IMG+BOW
• LSTM
• BOW

Model Comparison

• 2-VIS+BLSTM vs VIS+LSTM vs LSTM

Toronto COCO-QA

QAs: [link]

Images: Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar and C. Lawrence Zitnick, "Microsoft COCO: Common Objects in Context", ECCV 2014.

Notes:

1. All images are hosted on Flickr, and some links may not be available anymore.
2. Only test set results are rendered in the links below.

Individual Models

• 2-VIS+BLSTM
• VIS+LSTM
• IMG+BOW
• IMG+PRIOR
• IMG
• LSTM
• BOW

Model Comparison

• IMG+BOW vs 2-VIS+BLSTM vs IMG+PRIOR vs BOW