Posted: 10/27/2020 4:04 PM
(Edited: 10/27/2020 4:40 PM)
Post subject: Longplay movie files
Quote
Sanqui
Any
Player (25)
Joined: 4/25/2011
Posts: 33
Hi, I'm looking to use gameplay footage for machine learning purposes. There are a lot of longplays available on YouTube, but those universally suffer from compression artifacts stemming from video processing, plus I may want to use data from RAM to teach the neural network. TAS movie files are perfect for this purpose, because I can take original quality screenshots and log memory, but... TASes have the explicit goal of beating the game as fast as possible, while for machine learning you want as much footage of the game as possible, and those two goals are at odds. So my primary question is, does anybody know if there is a collection or database of movie files which are not TASes or speedruns? If not, I think it would be an interesting initiative to start.
ovo
Posted: 10/27/2020 4:16 PM
Quote
Spikestuff
They/Them
Editor, Publisher, Expert player (2312)
Joined: 10/12/2011
Posts: 6343
Location: The land down under.
Flash TASer of 2022 PSX TASer of 2016
Why not ask directly at the source? WoL for retro games uses movie files from tool-assisted emulators (for the majority of them) and there's a chance that some of the members have kept their files.
WebNations/Sabih wrote:
+fsvgm777 never censoring anything.
Disables Comments and Ratings for the YouTube account. Something better for yourself and also others.
Posted: 10/27/2020 4:44 PM
Quote
Sanqui
Any
Player (25)
Joined: 4/25/2011
Posts: 33
Spikestuff wrote:
Why not ask directly at the source? WoL for retro games uses movie files from tool-assisted emulators (for the majority of them) and there's a chance that some of the members have kept their files.
I actually didn't realize World of Longplays had such a nice database on their website. Unfortunately, it is just a database of videos, when I check out their Sonic 2 longplay the original movie file is not on offer. That's a shame, I suppose I should give them a visit and convince them that publishing those would be a good idea :)
ovo
Posted: 10/27/2020 6:55 PM
Quote
Aran_Jaeger
He/Him
Banned User, Player (9)
Joined: 10/29/2014
Posts: 176
Location: Bavaria, Germany
[quote Sanqui]So my primary question is, does anybody know if there is a collection or database of movie files which are not TASes or speedruns? If not, I think it would be an interesting initiative to start.[/quote] Actually I may be able to help you out on that, Sanqui. I have 100% completion gameplay SNES9X 1.53 movie files for around 400 to 500 Super Metroid (NTSC) hacks (for the SNES), so if that would be something that might help you, just contact me I guess. Someone that goes on Discord by the nickname barretrees had contacted me in 2018 about a very similar project on a by now published scientific paper (the PDF can be downloaded from here: https://dl.acm.org/doi/10.1145/3337722.3337728 ) of this kind before but in that case I think it was about automated ordering of gameplay moments based on certain RAM or visual information for the purpose of making a search engine for identifying and mapping search queries to moments in games, in my case mostly via video material (since I have corresponding videos for all of these movie files as well). Also, on the topic of completionist longplays (of broad varieties of games) that consist not solely of videos but are in the form of movie files for specified emulators, I have thought about the potential usefulness of that from a TASing perspective, too, because if one would as TASer be interested in a new game (either to make a TAS for it entirely or to test something out at some part of the game for which one has an idea, e.g. at parts late into the game) then it would be useful if a movie file would already exist (e.g. in the case that there is no existing TAS of the game yet) that one could fast-forward to quickly get to the point of interest.
collect, analyse, categorise. "Mathematics - When tool-assisted skills are just not enough" ;) Don't want to be taking up so much space adding to posts, but might be worth mentioning and letting others know for what games 1) already some TAS work has been done (ordered in decreasing amount, relative to a game completion) by me and 2) I am (in decreasing order) planning/considering to TAS them. Those would majorly be SNES games (if not, it will be indicated in the list) I'm focusing on. 1) Spanky's Quest; On the Ball/Cameltry; Musya; Super R-Type; Plok; Sutte Hakkun; The Wizard of Oz; Battletoads Doubledragon; Super Ghouls'n Ghosts; Firepower 2000; Brain Lord; Warios Woods; Super Turrican; The Humans. 2) Secret Command (SEGA); Star Force (NES); Hyperzone; Aladdin; R-Type 3; Power Blade 2 (NES); Super Turrican 2; First Samurai. (last updated: 18.03.2018)
Posted: 10/29/2020 6:40 PM
Quote
Sanqui
Any
Player (25)
Joined: 4/25/2011
Posts: 33
Aran Jaeger wrote:
Actually I may be able to help you out on that, Sanqui. I have 100% completion gameplay SNES9X 1.53 movie files for around 400 to 500 Super Metroid (NTSC) hacks (for the SNES), so if that would be something that might help you, just contact me I guess. Someone that goes on Discord by the nickname barretrees had contacted me in 2018 about a very similar project on a by now published scientific paper (the PDF can be downloaded from here: https://dl.acm.org/doi/10.1145/3337722.3337728 ) of this kind before but in that case I think it was about automated ordering of gameplay moments based on certain RAM or visual information for the purpose of making a search engine for identifying and mapping search queries to moments in games, in my case mostly via video material (since I have corresponding videos for all of these movie files as well). Also, on the topic of completionist longplays (of broad varieties of games) that consist not solely of videos but are in the form of movie files for specified emulators, I have thought about the potential usefulness of that from a TASing perspective, too, because if one would as TASer be interested in a new game (either to make a TAS for it entirely or to test something out at some part of the game for which one has an idea, e.g. at parts late into the game) then it would be useful if a movie file would already exist (e.g. in the case that there is no existing TAS of the game yet) that one could fast-forward to quickly get to the point of interest.
Aran, that is amazing information you have given me. The paper you've linked is super exciting and on a topic I've given much thought in the past, and certainly ties into some of my projects... I'll definitely be reading it and possibly getting in touch with the authors. I also agree that a database of longplay movie files would be a useful resource for TASing as well as speedrunners and others. With a smarter way to seek or at least some manual chapter checkpoints, it's can be used as a savestate database on steroids. Personally I'm not too interested in Super Metroid but I see huge value in your movie files of hacks from a preservationist standpoint. I think you should definitely publish them :)
ovo
Posted: 11/5/2020 7:38 PM
Quote
XTREMAL93
He/Him
Active player (389)
Joined: 1/6/2012
Posts: 579
Location: Azerbaijan, Baku
Hi. I keep my longplay movie files here. It is mostly NES, SNES and Sega Genesis games recorded in Bizhawk (bk2/tasproj). Feel free to use them for machine learning purposes. Also I would like to know more about your ML projects
Posted: 11/19/2020 1:30 PM
Quote
Sanqui
Any
Player (25)
Joined: 4/25/2011
Posts: 33
XTREMAL93 wrote:
Hi. I keep my longplay movie files here. It is mostly NES, SNES and Sega Genesis games recorded in Bizhawk (bk2/tasproj). Feel free to use them for machine learning purposes. Also I would like to know more about your ML projects
Hey XTREMAL93, thank you for your contribution! Slowly but surely, I'm building a collection. I finally managed to register on the World of Longplays forum, too (had problems with the verification email coming through), so I opened a thread there. As for my machine learning projects, so far I've trained a neural network to tell Sonic 1 and 2 screenshots apart, no matter how blurry or stretched they are. I've also built an auto-encoder for Sonic 1 screenshots, I describe how it works in this post. This is just the beginning though, I'm still learning. I have plenty of other ideas, like predicting actual data in memory or working with tiles as opposed to pixels for interesting visual output. However before I continue working on these projects I'll need multiple movie files of casual gameplay for each game I want to work with. YouTube playthroughs simply aren't enough.
ovo