Diverse Image Captioning via GroupTalk
Authors: Zhuhao Wang, Fei Wu, Weiming Lu, Jun Xiao, Xi Li, Zitong Zhang, Yueting Zhuang
IJCAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on several benchmark datasets show that Group Talk naturally diversifies the generated captions of each image without sacrificing the accuracy. |
| Researcher Affiliation | Academia | College of Computer Science, Zhejiang University, China {zhuhaow, wufei, luwm, junx, xilizju, ztzhang, yzhuang}@zju.edu.cn |
| Pseudocode | No | The paper describes algorithms verbally and through equations but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about the release of source code or links to a code repository. |
| Open Datasets | Yes | We run our experiments on Flickr8K [Hodosh et al., 2013], Flickr30K [Hessel et al., 2015] and MS COCO [Chen et al., 2015] which are three benchmark datasets widely used in image caption related tasks. |
| Dataset Splits | Yes | For Flickr8K and Flickr30K, we use 1, 000 images for validation and 1, 000 for testing, and for MS COCO, we use 5, 000 for validation and 5, 000 for testing. |
| Hardware Specification | No | The paper mentions training 'on CPU' and 'on GPU' but does not specify any particular models (e.g., specific CPU or GPU names, memory, etc.). |
| Software Dependencies | No | The paper mentions 'Python' and 'Lua with Torch based packages' but does not specify any version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We set the size of hidden units in the LSTM of describing models to 512 and classifier to 128. We use two describing models (Model 1/2) for Flickr8K and Flickr30K and three describing models (Model 1/2/3) for MS COCO. The number of describing model is set based on the size and the degree of diversity of the training data. Sentences are generated with beam size 5 (except for Base Model (S)). |