README.md
October 9, 2017 ยท View on GitHub
Generation of QAs
After playing Infinite Mario Bro. and recording event logs, now we can generate QAs.
Setup instructions
-
Install JDK, JRE and maven.
-
Compile the source code with following line.
$ bash run_compile.sh
Run to generate QAs
$ bash run_generation_QA.sh
- Inputs
data/logFileLists.txt: The filename list of event logs to generate QAs.data/configuration.json: Options for sampling clip region (fps, margin, durations).data/template.json: Pre-defined question template.data/phrase.json: Pre-defined phrase (or word) mapping to fill the question template or to obtain answer phrase.
- Outputs
generated_annotations/$gameplay_id$_raw_annotations.json: Raw annotation for each gameplay.generated_annotations/filtered_annotations_NT.json: Filtered out annotations with only NT questions.generated_annotations/filtered_annotations_ET.json: Filtered out annotations with only ET questions.generated_annotations/filtered_annotations_HT.json: Filtered out annotations with only HT questions.generated_annotations/filtered_annotations_ALL.json: Filtered out annotations with all questions.
- Annotation include following informations
qa_id: Question idvideo_path: The form isgameplay_id/gameplay_id_%d.datbegin_frame: The first frame number of video clipend_frame: The last frame number of video clipclip_length: The length of clipquestion: The question of clipanswer: The answer of clip for the questionevent: The target event in question, e.g.Kill,Eat,Hit, etcquestion_type: The question type -event-centric,counting,stateinterrogative: The interrogative of question, e.g.How,When,Who, etctemporal_relationship: Difficulty of temporal relationship -NT,ET,HTsemantic_chunk: Semantic chunk of qeustion, e.g.Event-Centric-Kill-How-Arg1-RedKoopa-Reference-After-Kill-Arg1-GreenKoopaWing. This is added since our ICCV 2017 paper, so sematic chunk is not included in MarioQA data used for ICCV 2017.
Download the event log files used in paper
$ bash run_download_event_logs.sh
Download the filtered annotations used in paper
$ bash run_download_filtered_annotation.sh