GameStory: An event-based Approach Van-Tu Ninh, Tu-Khiem Le, Minh-Triet Tran University of Science, VNU-HCM nvtu@apcs.vn,ltkhiem@apcs.vn,tmtriet@fit.hcmus.edu.vn ABSTRACT of game content and any other attached voices might overwhelm In this work, the authors propose an event-based approach to gen- the original game sound. erate summary of an e-sport match which satisfies certain criteria nominated by GameStory: The 2018 Video Game Analytics Chal- 2 APPROACH lenge. The authors’ solution aims to recite the match while empha- With the prior purpose of reciting appealing moments of the entire sizing the highlights of the game to keep the enthusiasm of viewers. match, the authors come up with an approach which is dividing the Our remedy analyses only player’s behaviour information logged game-play into several sequences and ranking them based on its from each match of the game to select important attributes as stan- excitement. Each extracted sequence is corresponding to a game- dard criteria for summary construction. From our perspective, the play event as recorded in the timelines’ structured data and belongs highlights of CS:GO game recorded in the summary should include to one of ten categories: assist, bomb_defuse, bomb_defuse_begin, killing scenes with headshot or penetrated, bomb planting, and bomb_plant, kill, purchase, round_end, round_start, suicide, and bomb defusing attempts. throw. According to the references on game summarization, we observe that the kill events and match progress are the cornerstone of 1 INTRODUCTION the game story and of the most appealing moments across the Video game analytics is a new research area which serves the need whole matches. Therefore, the authors focus on evaluating those of watching streaming e-sport games in a short time while capturing sequences to obtain the final weight in order to craft the demanding the essential content of these videos. Besides, analyzing video game synopsis through the sorted rank-list. also provides basic understanding of the gaming market including: what kind of games that the majority of viewers are interested in, 2.1 Pre-processing what event in the game makes the spectators excited, which part of There are various factors which could affect the entertainment the streaming video that viewers pay attention to most of the times, of the game story. The pre-processing focuses on analyzing the etc. However, video game summarization is a challenging task as extracted sequences under different criteria independently, and not only should the summary recite the full story of a long game in giving a corresponding score for each criterion. a short time (downscale from hours to minutes) but it should also make the spectators exciting with highlights emphasis, comments 2.1.1 Headshot. Obviously, this is one of the most important and sound. factors that defines a good kill in almost any shooting games. More- In GameStory: The 2018 Video Game Analytics Challenge, we over, the weapon which executes the kill also contributes to the are provided 11 CS:GO games which include game logs, game statis- amazement of the story. For instance, a head-shot caused by pistol tics, map overview, videos of player perspective, and commentator or sniper rifle is more appealing than the one caused by assault rifle. stream [1]. With these data, we propose a simple event-based solu- Therefore, we apply a score for each weapon in order to classify the tion to summarize CS:GO game with prominent events pre-defined kill and its excitement. The formula of the Headshot score would from our perspective. We analyse the game logs and select the be: weapon_score headshot . attributes that we consider as noticeable events in CS:GO game 2.1.2 Penetrating kills. Sudden death also entertains audiences which involves killing events with headshot or penetrated scenes, by its surprise. Penetrating kill is one cause of such deaths, which bomb planting and defusing actions. The idea of our proposed rem- is happened when the bullets of a player going through non-solid edy originates from a simple observation from CS:GO game which objects and hitting his enemy. is the more actions the players of both teams perform, the more thrilling the match is. From this idea, we statisticize, vectorize those 2.1.3 Consecutive kills and Multiple kills. Another factor of an attributes from game logs to create hand-crafted feature, and then intriguing match is the continuous kills of a player. The same is true multiplied them with our pre-defined weights to compute the score. for the one that has a lot of kills in a single round. Therefore, we Finally, we sorted the list according to the score and combine re- also take these two factors into account in generating the summary markable events from that ranked list to generate game summary. video. Although comments for game events could be attached to make 2.1.4 Round Bonus. It is also important to look at a general the summary more exciting, we do not utilize commentators’ voice manner which is how the game events linked with each other to or text due to the fact that the information shown in each player’s form an interesting story. From our observation, a round becomes perspective videos is enough to provide viewers with the process thrilling when a bomb is set by terrorists and two teams start to Copyright held by the owner/author(s). struggle to protect/defuse the bomb. Additionally, when it comes MediaEval’18, 29-31 October 2018, Sophia Antipolis, France to the decision round, the round itself is also extremely attractive. These explained why we construct a round bonus for each sequence MediaEval’18, 29-31 October 2018, Sophia Antipolis, France M. Larson et al. Table 1: Evaluation of HCMUS summary of CS:GO in 4 CONCLUSION GameStory: 2018 Video Game Analytics Challenge We present our proposed event-based solution to generate a sum- mary of CS:GO game streaming in GameStory: Video Game Analyt- HCMUS_Run1 ics Challenge based on event ranking. Our remedy is a simple and The submission gives a summary of the promising approach to apply not only in CS:GO game but also in 3.00 match at hand other game domains if game logs is detailed enough to create hand- The submission is entertaining 3.00 crafted features while creating competitive results of prominent The submission provides the flow and peak scenes. 3.67 of a good story The submission provided an innovative REFERENCES way to present a summary of an CS:GO 3.67 [1] Mathias Lux, Michael Riegler, Duc-Tien Dang-Nguyen, Marcus Lar- match son, Martin Potthast, and Pål Halvorsen. 2018. GameStory Task at A summary like this submission can be MediaEval 2018. In Working Notes Proceedings of the MediaEval 2018 2.67 Workshop. applied to games different from CS:GO so that we can emphasize the special rounds in the final result, and we calculate the bonus based on the bomb_plant, bomb_defuse, bomb_defuse_begin events. 2.2 Game Story Generation As our approach aims to tell the story by unifying game events, it is necessary to obtain the top exciting sequences corresponding to each event. The authors define a weight for each sequence which is calculated by the linear combination of the examining criteria as explained in 2.1 with sufficient coefficient to create a rank-list. The higher score the sequence obtains, the more thrilling it is. Based on the provided meta-data, we are able to split the players’ stream videos into smaller parts with duration of 3 seconds in length. Each part is corresponding to a sequence in the game-logs. With simple iteration through the rank-list, the authors are able to filter the top interesting sequences with adequate total length in duration. After that, we generate game summarization by combining these sequences in the ascending order of event’s happening time in order to preserve the flow of the story. 3 RESULTS AND ANALYSIS Table 1 displays the result of our approach with the evaluation from three different reviewers. According to tabel 1, we maintain entertainment of the game story at the average level of 3.00. This means that decision to focus on the kill and bomb involved events is appropriate. However, we think that we could modify our step by joining the small sequences into a summary video such as smooth- ing the transition among sequences, putting texts as well as player information, reporting the round status, and adding slow motion on particular kills. By doing so, the authors hope to obtain better score on story flow and peak compared to the 3.67 on our Run1 submission. With the event-based approach in summarizing game story, the authors achieve the goal of keeping the process simple while recit- ing the main events of the matches. The approach is dynamic as we could choose what story to tell actively just by modifying the coefficients in the linear combination of the examining criteria. This enables the capacity of our approach to be applied in other games and results in the score 2.67 in the cross-domain criteria on our submission.