Game of Sketches: Deep Recurrent Models of Pictionary-Style Word Guessing
Authors: Ravi Kiran Sarvadevabhatla, Shiv Surya, Trisha Mittal, R. Venkatesh Babu
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our model on the large-scale guess-word dataset generated via Sketch-QA task and compare with various baselines. We also conduct a Visual Turing Test to obtain human impressions of the guess-words generated by humans and our model. Experimental results demonstrate the promise of our approach for Pictionary and similarly themed games. |
| Researcher Affiliation | Academia | Ravi Kiran Sarvadevabhatla, Shiv Surya, Trisha Mittal, R. Venkatesh Babu Video Analytics Lab, Indian Institute of Science, Bangalore 560012, INDIA |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Please visit our project page http://val.cds.iisc.ac.in/sketchguess for supplementary material, code and dataset related to our work. |
| Open Datasets | Yes | Via Sketch-QA, we create a new crowdsourced dataset of paired guess-word and sketch-strokes, dubbed WORDGUESS-160, collected from 16,624 guess sequences of 1,108 subjects across 160 sketch object categories. Please visit our project page http://val.cds.iisc.ac.in/sketchguess for supplementary material, code and dataset related to our work. |
| Dataset Splits | Yes | For model evaluation, we split the 16,624 sequences in GUESSWORD-160 randomly into disjoint sets containing 60% , 25% and 15% of the data which are used during training, validation and testing phases respectively. |
| Hardware Specification | No | The paper does not provide specific hardware details used for running its experiments. |
| Software Dependencies | No | The paper mentions several software components like 'Hun Pos tagger', 'Enchant spell check library', 'word2vec', 'VGG-16', 'LSTM', and 'Adagrad optimizer', but it does not specify their version numbers. |
| Experiment Setup | Yes | For all the experiments, we use Adagrad optimizer (Duchi, Hazan, and Singer 2011) with a starting learning rate of 0.01 and early-stopping as the criterion for terminating optimization. The value for margin is set to 0.1. Overall, we found the convex combination loss with λ = 1 (determined via grid search) to provide the best performance. LSTM with 512 hidden units as the RNN component. |