Generating an Event Timeline About Daily Activities From a Semantic Concept Stream
Authors: Taiki Miyanishi, Jun-ichiro Hirayama, Takuya Maekawa, Motoaki Kawanabe
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The empirical evaluation demonstrates that our method using real-world properties improves the performance of generating an event timeline over diverse environments. |
| Researcher Affiliation | Academia | Taiki Miyanishi,1 Jun-ichiro Hirayama,2,1 Takuya Maekawa,3,1 Motoaki Kawanabe1,2 1Advanced Telecommunications Research Institute International (ATR), Kyoto, Japan 2RIKEN Center for Advanced Intelligence Project (AIP), Tokyo, Japan 3Graduate School of Information Science and Technology, Osaka University, Osaka, Japan |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures). |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described in this paper; there is no specific repository link, explicit code release statement, or code in supplementary materials. |
| Open Datasets | No | The paper mentions using a 'daily-activities dataset with manually labeled event descriptions' which they annotated, and also states, 'This dataset has been used in egocentric video retrieval with gesture motions (Miyanishi et al. 2016)'. However, no specific link, DOI, repository name, or explicit statement of public availability for this dataset is provided. |
| Dataset Splits | Yes | Then, we predicted labels of semantic concepts with leave-one-cross-session training, which trains a model on the nine sessions, and tested it with another session s data. ... Our training set for the language model was a total of 51,564 events, and 6,871 events were used for validation. ... We used leave-one-place-out cross-validation, optimized for the best performance in Bleu score (as described later), on the validation data of five places (not including a target place). |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like GRU, Adam optimizer, Glo Ve, Stanford dependency parser, and VGG, but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | We set the window size to N=3 sec. We trained our GRU using the Adam optimizer (Kingma and Ba 2015), with a learning rate of 0.001 and a batch size of 32. We set hidden dimensions of GRU 300 initialized with a Normal distribution N(0, 0.01). Its parameters were optimized to get the best validation performance in training runs for up to 30 epochs. We used the cross entropy loss for a multi-class label of action and place. We used a multi-label one-versus-all loss based on the max-entropy for multi-labeling of object labels. ... We set hidden dimensions of Glove 300 following (Pennington, Socher, and Manning 2014). ... We tuned a few hyper-parameters the window-size M for generating candidates of the event, α for the temporal model, and β for the event sequence model among candidates M = {1, 2, 3, 4, 5}, α = {0, 0.2, 0.4, 0.6, 0.8} and β = {0, 0.2, 0.4, 0.6, 0.8}. |