Sep 24, 2021

Computer Vision – a technology behind ELAB’s ward heat maps

Computer Vision – a technology behind ELAB’s ward heat maps

The third technology we want to introduce to you is Computer Vision. With this breakthrough technology, we can analyze an entire League of Legends match in under a minute and provide layers of data that were previously unavailable.


Technical presentation of ELAB Computer Vision technology 

In ELAB, we have designed and built an AI-fueled pipeline for automatic statistics generation from League of Legends game video replays. Let us first address the obvious question – why didn’t we just use the publicly available RIOT API to gather data? The data available in API is restricted to SoloQ games, and moreover, the game-states are provided with one-minute ticks. This is obviously restrictive from data-analytic perspective. Unfortunately, extracting game-state information directly from LoL client replay file .rofl is also being restricted by RIOT ToS.  

Our AI pipeline is capable of translating every single frame of any LoL game replay into a tabular data structure comprehensively capturing game-state including champion location, stats, gold, items, turrets, wards and even minion locations. Our pipeline is based upon advanced computer vision technologies. The main building block of our pipeline is numerous instances of artificial neural networks. We employed modern architectures including YOLO & CenterNet.

Simplifying a bit, those convolutional neural nets architectures are inspired by the brain, i.e. neurons (perceptrons) are stacked up in quite a few layers, where each subsequent layer is responsible for the recognition of coarser features.

However, to be able to perform detection and classification those nets need to be appropriately tuned and trained at first. The standard applications of those nets that you may have seen in the literature deal with detecting objects on natural images, i.e. photos.

In our application we had to introduce many tweaks to be able to deal with synthetic images like the ones rendered by LoL game engine. We will be revealing more details about our technology and approach to data mining of extracted LoL game states in an upcoming scientific publication, so make sure to follow our posts on social media. 

Having the game-states extracted from replays as the next step we can extract numerous statistics. One prominent example presented below – the ward heatmaps, which reveal how the selected team handles in-game vision. We are working intensely on many more stats and will be revealing soon jungle pathing, and champion death heatmaps. Soon, we will be creating stats, which use all the computer vision aggregated stats at once, for example champion group efficiency.


EsportsLab x Ultraliga

If you follow our social media regularly, you may have seen the first capabilities of Computer Vision technology back at this year's Mid-Season Invitational. It was then that the first ward heat maps appeared on our Twitter and Facebook. We took the last 10 MAD Lions games and combined them into two different heat maps, for the blue and red side of Summoner's Rift.


A month later, another important step in the development of our technology took place. On exactly July 7 we announced to the world our partnership with Ultraliga, Poland's most prestigious League of Legends competition. For the last few weeks, you've been able to see our heat maps during the Ultraliga broadcast. Our data has already enriched the analysis of the match between Illuminar Gaming and K1CK or the already historic match between PDW and AGO ROGUE. An example of heat maps that regularly appear during the broadcast can be seen below.


Here we have chosen the semifinal of Mid-Season Invitational 2021, RNG - PSG. RNG defeated the PCS representative by a score of 3:1 and advanced to the grand final, where they took the title after a win against DK. With our heat maps, you can clearly see the trends in RNG's game, which were one of the reasons for their success at the international tournament.

In the match against PSG, bot lane was in the spotlight. Gala, known for his great Kai'Sa, and Ming, considered one of the best supporters in the world at the time, were in RNG's eyes their easiest path to victory. Regardless of the side of the map, RNG had a very strong vision at the bottom of the map. They wanted to win bot lane and very often it worked out for them. On the top of the map, the weak side player Xiaohu was supported by a defensive vision that prevented cross-map plays. The vision ensured that Xiaohu was safe and could retreat at the right time when he saw his opponents rotating.

With the knowledge gained from heat maps, we are even able to prepare a strategy against RNG in this encounter. The main concern is of course bot lane, which could be countered with an aggressive early game duo and oppressive jungler. This way we stop Gala from safely farming and don't allow him to quickly reach a state where he can almost single-handedly win the game.

In another scenario, we can bet on a more defensive bot lane that scales well to the late game. A lot then depends on the skill of our players, of course. They will probably spend most of the early game under the turret and will have to respond to attacks from Gala, Ming, and Wei. For the top lane, we can choose a Camille-style champion, who is a meta character, scales well, and can keep up with our bot lane in the late game.

Heat maps are therefore not only a tool to enrich the broadcast and support experts during analysis, but also an excellent method for coaches to understand their opponents. In the coming months, you will see even more of our heat maps, and soon you will also learn about the rest of the Computer Vision technology.


Got questions?

Contact one of our departments to solve your problem.


Are you an owner of an esports team or maybe want to develop your division? Feel free to contact us!

General questions

Any other questions? Don’t hesitate to contact us!