“This is Home Invasion,” Korine says, pointing at a computer monitor inside the house. On the screen plays security--camera and GoPro footage of masked men and women, wearing horns, rampaging through various Miami McMansions, tying up various innocent-looking families. “We’re trying to gamify movies,” Korine says. So Home Invasion is designed to look, at times, like a first-per...
(展开全部)
“This is Home Invasion,” Korine says, pointing at a computer monitor inside the house. On the screen plays security--camera and GoPro footage of masked men and women, wearing horns, rampaging through various Miami McMansions, tying up various innocent-looking families. “We’re trying to gamify movies,” Korine says. So Home Invasion is designed to look, at times, like a first-person shooter, and inside the film the home invaders are sometimes playing games —the idea being you can scan a QR code and play along with them. “What we’re trying to do is to build some mechanism that allows people to interface with the footage and basically remix, or make their own, films,” Korine says.
电影游戏化，现实世界的变脸抢劫RPG。nn首先，有三个明确主体，侵略者、导播、富豪。nn电影展示了两个大方向的技术，目标检测（object detection）和图像处理（image processing）。前者主要用于识别人类目标，后者对检测到的目标进行处理。nn考虑到直播的实时性，检测模型应该用单阶段检测器，延迟小，像最新版本的yolo这种，应该比较适配。同时，侵略人员和富豪数量有限，不存在检测目标过多过密的情况，检测精度应该挺高的。电影呈现的目标检测效果主要有两种，一种是人物的五官识别，一种是人物的动作识别，这两项技术都比较成熟，有大量开源的训练好的模型可以使用。nn针对侵略者，后期需要保护其隐私性，因此需要进行识别后结果再处理，而富豪均直接展示五官和姿态结果，未进行再一步处理。猜测可能对侵略者做了数据图像存储处理，在监测到人物五官后，与数据库中人员图像进行比对，如果一致则需要进行再处理。电影中最直接的方法是局部模糊化，这里可以用高斯函数之类的方法。其次，就是baby invader化处理，将人物脸部区域替换成婴儿，电影中应该接入了AI，调用api直接处理。（但电影中婴儿脸形式太单一了，可以用个lora训练一类稳定有特点的，不然看着不稳定又有点视觉疲劳）nn任务提示板是导播的任务，对话型文字内容可考虑语音识别，任务提示型文字应该是方便观众观看，同时增加互动性。风格化场景、一些莫须有的动物、金币表情等也是直接利用AI，这部分元素比较无感，组合比较乱。nn电影形式新颖，但有点没放开，我想看侵略者牵手唱儿歌，突突突流血场景啊！btw音乐是Burial做的。nn我乱说完了，大家去试试吧。