Repository navigation
Expand file tree
/
Copy pathtimeline.js
More file actions
1 lines (1 loc) · 60.7 KB
/
Copy pathtimeline.js
File metadata and controls
1 lines (1 loc) · 60.7 KB
1
window.TL={"title": "压缩三十年:从 H.263 到 H.266 / AV2 / AVS3", "subtitle": "主流视频编码格式科普 · I/P/B · 率失真 · 档次与微调", "duration": 956.735, "segments": [{"id": "s01", "ch": "开场", "scene": "hook", "sfx": ["random/levelup"], "text": "一段一分钟的1080p、60帧视频,如果一个像素都不压缩,需要多大?每帧三百多万字节,每秒将近1.5G比特,一分钟就是十一个G。可你在网上看它,只用了五兆比特每秒。压缩了三百倍,肉眼却几乎看不出差别。这三百倍,是三十年里一代又一代视频编码标准,一点一点抠出来的。", "start": 0.6, "dur": 26.233, "voice": "voice/s01.wav", "vdur": 25.033375, "sents": [{"t0": 0.35, "t1": 6.59, "text": "一段一分钟的1080p、60帧视频,如果一个像素都不压缩,需要多大?"}, {"t0": 6.59, "t1": 12.57, "text": "每帧三百多万字节,每秒将近1.5G比特,一分钟就是十一个G。"}, {"t0": 12.57, "t1": 15.99, "text": "可你在网上看它,只用了五兆比特每秒。"}, {"t0": 15.99, "t1": 19.75, "text": "压缩了三百倍,肉眼却几乎看不出差别。"}, {"t0": 19.75, "t1": 25.383, "text": "这三百倍,是三十年里一代又一代视频编码标准,一点一点抠出来的。"}], "subs": [{"t0": 0.35, "t1": 3.29, "text": "一段一分钟的1080p、60帧视频"}, {"t0": 3.31, "t1": 6.57, "text": "如果一个像素都不压缩,需要多大?"}, {"t0": 6.59, "t1": 10.63, "text": "每帧三百多万字节,每秒将近1.5G比特"}, {"t0": 10.65, "t1": 12.55, "text": "一分钟就是十一个G"}, {"t0": 12.57, "t1": 15.97, "text": "可你在网上看它,只用了五兆比特每秒"}, {"t0": 15.99, "t1": 19.73, "text": "压缩了三百倍,肉眼却几乎看不出差别"}, {"t0": 19.75, "t1": 23.65, "text": "这三百倍,是三十年里一代又一代视频编码标准"}, {"t0": 23.67, "t1": 25.383, "text": "一点一点抠出来的"}]}, {"id": "s02", "ch": "开场", "scene": "title", "rec": [[120, 150]], "sfx": ["ui/toast/challenge_complete"], "text": "这期视频,我们借着一个开源的帧级分析工具VideoAnalyzer-Qt,从H.263、VP6、Theora,一路走到AVS3、H.266和AV2,把主流视频编码格式拆开来看:它们解决了什么痛点,背后的原理是什么,基线怎么选,参数又该怎么调。", "start": 26.833, "dur": 21.233, "voice": "voice/s02.wav", "vdur": 20.033375, "sents": [{"t0": 0.35, "t1": 20.383, "text": "这期视频,我们借着一个开源的帧级分析工具VideoAnalyzer-Qt,从H.263、VP6、Theora,一路走到AVS3、H.266和AV2,把主流视频编码格式拆开来看:它们解决了什么痛点,背后的原理是什么,基线怎么选,参数又该怎么调。"}], "subs": [{"t0": 0.35, "t1": 1.37, "text": "这期视频"}, {"t0": 1.39, "t1": 5.35, "text": "我们借着一个开源的帧级分析工具VideoAnal"}, {"t0": 5.37, "t1": 5.67, "text": "yzer-Qt"}, {"t0": 5.37, "t1": 7.65, "text": "从H.263、VP6、Theora"}, {"t0": 7.67, "t1": 10.73, "text": "一路走到AVS3、H.266和AV2"}, {"t0": 10.75, "t1": 15.19, "text": "把主流视频编码格式拆开来看:它们解决了什么痛点"}, {"t0": 15.21, "t1": 20.383, "text": "背后的原理是什么,基线怎么选,参数又该怎么调"}]}, {"id": "s03", "ch": "一 · 原理", "scene": "redundancy", "sfx": ["item/book/open_flip1"], "text": "视频能被压缩,是因为它充满了冗余。第一种是空间冗余:一块蓝天,相邻像素几乎一样。第二种是时间冗余:相邻两帧,大部分画面只是挪了一点位置。第三种是统计冗余:有些符号出现得特别频繁。最后还有感知冗余:人眼对颜色细节和高频纹理,远没有对亮度轮廓那么敏感。", "start": 48.067, "dur": 25.313, "voice": "voice/s03.wav", "vdur": 24.113375, "sents": [{"t0": 0.35, "t1": 3.49, "text": "视频能被压缩,是因为它充满了冗余。"}, {"t0": 3.49, "t1": 8.33, "text": "第一种是空间冗余:一块蓝天,相邻像素几乎一样。"}, {"t0": 8.33, "t1": 13.77, "text": "第二种是时间冗余:相邻两帧,大部分画面只是挪了一点位置。"}, {"t0": 13.77, "t1": 17.91, "text": "第三种是统计冗余:有些符号出现得特别频繁。"}, {"t0": 17.91, "t1": 24.463, "text": "最后还有感知冗余:人眼对颜色细节和高频纹理,远没有对亮度轮廓那么敏感。"}], "subs": [{"t0": 0.35, "t1": 3.47, "text": "视频能被压缩,是因为它充满了冗余"}, {"t0": 3.49, "t1": 8.31, "text": "第一种是空间冗余:一块蓝天,相邻像素几乎一样"}, {"t0": 8.33, "t1": 11.13, "text": "第二种是时间冗余:相邻两帧"}, {"t0": 11.15, "t1": 13.75, "text": "大部分画面只是挪了一点位置"}, {"t0": 13.77, "t1": 17.89, "text": "第三种是统计冗余:有些符号出现得特别频繁"}, {"t0": 17.91, "t1": 22.17, "text": "最后还有感知冗余:人眼对颜色细节和高频纹理"}, {"t0": 22.19, "t1": 24.463, "text": "远没有对亮度轮廓那么敏感"}]}, {"id": "s04", "ch": "一 · 原理", "scene": "hybrid", "sfx": ["block/beacon/power1"], "text": "几乎所有主流编码器,都建立在同一套框架上,叫做基于块的混合编码。画面先被切成块。每个块先做预测:要么参考同一帧里的邻居,叫帧内预测;要么参考别的帧,叫帧间预测。预测不准的残差,再经过变换、量化,最后熵编码成比特。解码端把这一切倒过来做一遍。", "start": 73.38, "dur": 25.003, "voice": "voice/s04.wav", "vdur": 23.803375, "sents": [{"t0": 0.35, "t1": 6.63, "text": "几乎所有主流编码器,都建立在同一套框架上,叫做基于块的混合编码。"}, {"t0": 6.63, "t1": 8.49, "text": "画面先被切成块。"}, {"t0": 8.49, "t1": 13.39, "text": "每个块先做预测:要么参考同一帧里的邻居,叫帧内预测;"}, {"t0": 13.39, "t1": 16.63, "text": "要么参考别的帧,叫帧间预测。"}, {"t0": 16.63, "t1": 21.65, "text": "预测不准的残差,再经过变换、量化,最后熵编码成比特。"}, {"t0": 21.65, "t1": 24.153, "text": "解码端把这一切倒过来做一遍。"}], "subs": [{"t0": 0.35, "t1": 4.05, "text": "几乎所有主流编码器,都建立在同一套框架上"}, {"t0": 4.07, "t1": 6.61, "text": "叫做基于块的混合编码"}, {"t0": 6.63, "t1": 8.47, "text": "画面先被切成块"}, {"t0": 8.49, "t1": 12.13, "text": "每个块先做预测:要么参考同一帧里的邻居"}, {"t0": 12.15, "t1": 13.37, "text": "叫帧内预测"}, {"t0": 13.39, "t1": 16.61, "text": "要么参考别的帧,叫帧间预测"}, {"t0": 16.63, "t1": 19.77, "text": "预测不准的残差,再经过变换、量化"}, {"t0": 19.79, "t1": 21.63, "text": "最后熵编码成比特"}, {"t0": 21.65, "t1": 24.153, "text": "解码端把这一切倒过来做一遍"}]}, {"id": "s05", "ch": "一 · 原理", "scene": "dct", "sfx": ["block/amethyst/resonate1"], "text": "变换这一步的核心,是1974年Ahmed等人提出的离散余弦变换。它把像素块转换成频率系数,能量会集中到左上角少数几个低频系数里。量化再把人眼不敏感的高频系数,大胆地削成零。整个压缩流程里,真正丢掉信息的,只有量化这一步。", "start": 98.384, "dur": 22.243, "voice": "voice/s05.wav", "vdur": 21.043375, "sents": [{"t0": 0.35, "t1": 5.63, "text": "变换这一步的核心,是1974年Ahmed等人提出的离散余弦变换。"}, {"t0": 5.63, "t1": 12.29, "text": "它把像素块转换成频率系数,能量会集中到左上角少数几个低频系数里。"}, {"t0": 12.29, "t1": 16.71, "text": "量化再把人眼不敏感的高频系数,大胆地削成零。"}, {"t0": 16.71, "t1": 21.393, "text": "整个压缩流程里,真正丢掉信息的,只有量化这一步。"}], "subs": [{"t0": 0.35, "t1": 2.09, "text": "变换这一步的核心"}, {"t0": 2.11, "t1": 5.61, "text": "是1974年Ahmed等人提出的离散余弦变换"}, {"t0": 5.63, "t1": 8.33, "text": "它把像素块转换成频率系数"}, {"t0": 8.35, "t1": 12.27, "text": "能量会集中到左上角少数几个低频系数里"}, {"t0": 12.29, "t1": 16.69, "text": "量化再把人眼不敏感的高频系数,大胆地削成零"}, {"t0": 16.71, "t1": 21.393, "text": "整个压缩流程里,真正丢掉信息的,只有量化这一步"}]}, {"id": "s06", "ch": "一 · 原理", "scene": "rdo", "sfx": ["block/enchantment_table/enchant1"], "say": "那么编码器怎么在几十上百种模式里做选择?答案是率失真优化。Sullivan和Wiegand在1998年的经典论文,把它写成一个式子:J等于D加lambda乘R。D是失真,R是码率,lambda是两者之间的汇率。每一个块,编码器都在寻找让J最小的那个选择。后面你听到的所有参数,本质上都在调这个汇率,和搜索的范围。", "text": "那么编码器怎么在几十上百种模式里做选择?答案是率失真优化。Sullivan和Wiegand在1998年的经典论文,把它写成一个式子:J = D + λ·R。D是失真,R是码率,λ是两者之间的汇率。每一个块,编码器都在寻找让J最小的那个选择。后面你听到的所有参数,本质上都在调这个汇率,和搜索的范围。", "start": 120.627, "dur": 29.653, "voice": "voice/s06.wav", "vdur": 28.453375, "sents": [{"t0": 0.35, "t1": 4.17, "text": "那么编码器怎么在几十上百种模式里做选择?"}, {"t0": 4.17, "t1": 6.37, "text": "答案是率失真优化。"}, {"t0": 6.37, "t1": 14.05, "text": "Sullivan和Wiegand在1998年的经典论文,把它写成一个式子:J = D + λ·R。"}, {"t0": 14.05, "t1": 18.63, "text": "D是失真,R是码率,λ是两者之间的汇率。"}, {"t0": 18.63, "t1": 22.576, "text": "每一个块,编码器都在寻找让J最小的那个选择。"}, {"t0": 22.576, "t1": 28.803, "text": "后面你听到的所有参数,本质上都在调这个汇率,和搜索的范围。"}], "subs": [{"t0": 0.35, "t1": 4.15, "text": "那么编码器怎么在几十上百种模式里做选择?"}, {"t0": 4.17, "t1": 6.35, "text": "答案是率失真优化"}, {"t0": 6.37, "t1": 10.369, "text": "Sullivan和Wiegand在1998年的经典论文"}, {"t0": 10.389, "t1": 14.03, "text": "把它写成一个式子:J = D + λ·R"}, {"t0": 14.05, "t1": 18.61, "text": "D是失真,R是码率,λ是两者之间的汇率"}, {"t0": 18.63, "t1": 22.556, "text": "每一个块,编码器都在寻找让J最小的那个选择"}, {"t0": 22.576, "t1": 27.038, "text": "后面你听到的所有参数,本质上都在调这个汇率"}, {"t0": 27.058, "t1": 28.803, "text": "和搜索的范围"}]}, {"id": "s07", "ch": "二 · I/P/B", "scene": "ipb_i", "sfx": ["note/pling"], "text": "接下来是最重要的概念:I帧、P帧和B帧。I帧是帧内编码帧,只用自己这一帧的信息,可以独立解码,是随机访问和拖动进度条的落脚点,代价是体积最大。", "start": 150.28, "dur": 15.547, "voice": "voice/s07.wav", "vdur": 14.346688, "sents": [{"t0": 0.35, "t1": 3.99, "text": "接下来是最重要的概念:I帧、P帧和B帧。"}, {"t0": 3.99, "t1": 14.697, "text": "I帧是帧内编码帧,只用自己这一帧的信息,可以独立解码,是随机访问和拖动进度条的落脚点,代价是体积最大。"}], "subs": [{"t0": 0.35, "t1": 3.97, "text": "接下来是最重要的概念:I帧、P帧和B帧"}, {"t0": 3.99, "t1": 8.47, "text": "I帧是帧内编码帧,只用自己这一帧的信息"}, {"t0": 8.49, "t1": 13.07, "text": "可以独立解码,是随机访问和拖动进度条的落脚点"}, {"t0": 13.09, "t1": 14.697, "text": "代价是体积最大"}]}, {"id": "s08", "ch": "二 · I/P/B", "scene": "ipb_p", "sfx": ["note/bell"], "text": "P帧是前向预测帧。它参考之前已经解码的帧,只记录运动矢量和残差。画面里一辆车往右开了十个像素,P帧只需要说一句:这块,从上一帧左边十个像素那里搬过来。所以P帧通常只有I帧的几分之一。", "start": 165.827, "dur": 19.747, "voice": "voice/s08.wav", "vdur": 18.546688, "sents": [{"t0": 0.35, "t1": 2.29, "text": "P帧是前向预测帧。"}, {"t0": 2.29, "t1": 7.71, "text": "它参考之前已经解码的帧,只记录运动矢量和残差。"}, {"t0": 7.71, "t1": 16.01, "text": "画面里一辆车往右开了十个像素,P帧只需要说一句:这块,从上一帧左边十个像素那里搬过来。"}, {"t0": 16.01, "t1": 18.897, "text": "所以P帧通常只有I帧的几分之一。"}], "subs": [{"t0": 0.35, "t1": 2.27, "text": "P帧是前向预测帧"}, {"t0": 2.29, "t1": 7.69, "text": "它参考之前已经解码的帧,只记录运动矢量和残差"}, {"t0": 7.71, "t1": 12.47, "text": "画面里一辆车往右开了十个像素,P帧只需要说一句"}, {"t0": 12.49, "t1": 15.99, "text": "这块,从上一帧左边十个像素那里搬过来"}, {"t0": 16.01, "t1": 18.897, "text": "所以P帧通常只有I帧的几分之一"}]}, {"id": "s09", "ch": "二 · I/P/B", "scene": "ipb_b", "sfx": ["note/icechime"], "text": "B帧是双向预测帧,它既能看过去,也能看未来,把前后两帧的预测加权平均。被遮挡又露出来的背景,在未来帧里能找到;噪声在两次预测的平均中也会互相抵消。所以B帧最小,往往只有I帧的五分之一甚至更少。", "start": 185.574, "dur": 20.447, "voice": "voice/s09.wav", "vdur": 19.246688, "sents": [{"t0": 0.35, "t1": 7.87, "text": "B帧是双向预测帧,它既能看过去,也能看未来,把前后两帧的预测加权平均。"}, {"t0": 7.87, "t1": 11.97, "text": "被遮挡又露出来的背景,在未来帧里能找到;"}, {"t0": 11.97, "t1": 15.29, "text": "噪声在两次预测的平均中也会互相抵消。"}, {"t0": 15.29, "t1": 19.597, "text": "所以B帧最小,往往只有I帧的五分之一甚至更少。"}], "subs": [{"t0": 0.35, "t1": 4.93, "text": "B帧是双向预测帧,它既能看过去,也能看未来"}, {"t0": 4.95, "t1": 7.85, "text": "把前后两帧的预测加权平均"}, {"t0": 7.87, "t1": 11.95, "text": "被遮挡又露出来的背景,在未来帧里能找到"}, {"t0": 11.97, "t1": 15.27, "text": "噪声在两次预测的平均中也会互相抵消"}, {"t0": 15.29, "t1": 19.597, "text": "所以B帧最小,往往只有I帧的五分之一甚至更少"}]}, {"id": "s10", "ch": "二 · I/P/B", "scene": "ipb_reorder", "sfx": ["ui/loom/select_pattern1"], "text": "但B帧要参考未来,就意味着解码顺序和显示顺序不再一致。编码器必须先把后面的P帧发出去,再发中间的B帧。在分析工具里,这就是最下面那条pts减dts的重排曲线:有B帧就有锯齿,没有B帧就是一条平线。代价是延迟,所以视频通话和直播的低延迟模式,往往会关掉B帧。", "start": 206.02, "dur": 25.943, "voice": "voice/s10.wav", "vdur": 24.743375, "sents": [{"t0": 0.35, "t1": 5.43, "text": "但B帧要参考未来,就意味着解码顺序和显示顺序不再一致。"}, {"t0": 5.43, "t1": 10.05, "text": "编码器必须先把后面的P帧发出去,再发中间的B帧。"}, {"t0": 10.05, "t1": 19.51, "text": "在分析工具里,这就是最下面那条pts减dts的重排曲线:有B帧就有锯齿,没有B帧就是一条平线。"}, {"t0": 19.51, "t1": 25.093, "text": "代价是延迟,所以视频通话和直播的低延迟模式,往往会关掉B帧。"}], "subs": [{"t0": 0.35, "t1": 2.03, "text": "但B帧要参考未来"}, {"t0": 2.05, "t1": 5.41, "text": "就意味着解码顺序和显示顺序不再一致"}, {"t0": 5.43, "t1": 10.03, "text": "编码器必须先把后面的P帧发出去,再发中间的B帧"}, {"t0": 10.05, "t1": 11.63, "text": "在分析工具里"}, {"t0": 11.65, "t1": 15.65, "text": "这就是最下面那条pts减dts的重排曲线"}, {"t0": 15.67, "t1": 19.49, "text": "有B帧就有锯齿,没有B帧就是一条平线"}, {"t0": 19.51, "t1": 23.53, "text": "代价是延迟,所以视频通话和直播的低延迟模式"}, {"t0": 23.55, "t1": 25.093, "text": "往往会关掉B帧"}]}, {"id": "s11", "ch": "二 · I/P/B", "scene": "rec", "rec": [[668, 700]], "sfx": ["ui/toast/in"], "text": "从一个I帧到下一个I帧,叫一个GOP。看这段MPEG-2的片子:红色是I,蓝色是P,绿色是B,典型的IBBP十二帧周期,码率图上是一座高峰、几个半山腰、一片低谷的反复。后来Schwarz等人在2006年分析了分层B帧结构:让B帧也被别的B帧参考,按金字塔层级分配量化参数。这成了H.264之后所有编码器的标配。", "start": 231.964, "dur": 30.383, "voice": "voice/s11.wav", "vdur": 29.183375, "sents": [{"t0": 0.35, "t1": 3.93, "text": "从一个I帧到下一个I帧,叫一个GOP。"}, {"t0": 3.93, "t1": 16.43, "text": "看这段MPEG-2的片子:红色是I,蓝色是P,绿色是B,典型的IBBP十二帧周期,码率图上是一座高峰、几个半山腰、一片低谷的反复。"}, {"t0": 16.43, "t1": 26.29, "text": "后来Schwarz等人在2006年分析了分层B帧结构:让B帧也被别的B帧参考,按金字塔层级分配量化参数。"}, {"t0": 26.29, "t1": 29.533, "text": "这成了H.264之后所有编码器的标配。"}], "subs": [{"t0": 0.35, "t1": 3.91, "text": "从一个I帧到下一个I帧,叫一个GOP"}, {"t0": 3.93, "t1": 8.17, "text": "看这段MPEG-2的片子:红色是I,蓝色是P"}, {"t0": 8.19, "t1": 12.03, "text": "绿色是B,典型的IBBP十二帧周期"}, {"t0": 12.05, "t1": 16.41, "text": "码率图上是一座高峰、几个半山腰、一片低谷的反复"}, {"t0": 16.43, "t1": 20.77, "text": "后来Schwarz等人在2006年分析了分层B帧结构"}, {"t0": 20.79, "t1": 23.47, "text": "让B帧也被别的B帧参考"}, {"t0": 23.49, "t1": 26.27, "text": "按金字塔层级分配量化参数"}, {"t0": 26.29, "t1": 29.533, "text": "这成了H.264之后所有编码器的标配"}]}, {"id": "s12", "ch": "三 · 演进", "scene": "timeline", "sfx": ["fireworks/launch1", "fireworks/twinkle1"], "text": "有了这些概念,我们沿着时间线走一遍。这三十年,基本是两条路线在赛跑:一条是ITU和MPEG的国际标准,从H.261、MPEG-2、H.264,到H.265、H.266;另一条是On2、谷歌和开放媒体联盟的开放格式,从VP3、VP6、Theora,到VP8、VP9、AV1、AV2。而中国的AVS标准,走出了第三条路。", "start": 262.347, "dur": 27.803, "voice": "voice/s12.wav", "vdur": 26.603375, "sents": [{"t0": 0.35, "t1": 4.01, "text": "有了这些概念,我们沿着时间线走一遍。"}, {"t0": 4.01, "t1": 14.91, "text": "这三十年,基本是两条路线在赛跑:一条是ITU和MPEG的国际标准,从H.261、MPEG-2、H.264,到H.265、H.266;"}, {"t0": 14.91, "t1": 23.59, "text": "另一条是On2、谷歌和开放媒体联盟的开放格式,从VP3、VP6、Theora,到VP8、VP9、AV1、AV2。"}, {"t0": 23.59, "t1": 26.953, "text": "而中国的AVS标准,走出了第三条路。"}], "subs": [{"t0": 0.35, "t1": 3.99, "text": "有了这些概念,我们沿着时间线走一遍"}, {"t0": 4.01, "t1": 7.33, "text": "这三十年,基本是两条路线在赛跑"}, {"t0": 7.35, "t1": 10.13, "text": "一条是ITU和MPEG的国际标准,从H.261"}, {"t0": 10.15, "t1": 13.212, "text": "MPEG-2、H.264,到H.265"}, {"t0": 13.232, "t1": 14.89, "text": "H.266"}, {"t0": 14.91, "t1": 18.85, "text": "另一条是On2、谷歌和开放媒体联盟的开放格式"}, {"t0": 18.87, "t1": 21.415, "text": "从VP3、VP6、Theora,到VP8"}, {"t0": 21.435, "t1": 23.57, "text": "VP9、AV1、AV2"}, {"t0": 23.59, "t1": 26.953, "text": "而中国的AVS标准,走出了第三条路"}]}, {"id": "s13", "ch": "H.263 · 1996", "scene": "codec", "codec": "h263", "rec": [[176, 192], [584, 600]], "sfx": ["ui/cartography_table/drawmap1"], "text": "1996年的H.263,痛点是电话线和早期移动网络上的可视电话:几十k的带宽,延迟就是一切。它的基线档只有I帧和P帧,半像素运动矢量,而且只支持几种固定尺寸,从128乘96,到CIF的352乘288。为什么不要B帧?因为B帧要等未来帧,每多等一帧,对通话都是不可接受的。", "start": 290.15, "dur": 28.563, "voice": "voice/s13.wav", "vdur": 27.363375, "sents": [{"t0": 0.35, "t1": 9.29, "text": "1996年的H.263,痛点是电话线和早期移动网络上的可视电话:几十k的带宽,延迟就是一切。"}, {"t0": 9.29, "t1": 19.783, "text": "它的基线档只有I帧和P帧,半像素运动矢量,而且只支持几种固定尺寸,从128乘96,到CIF的352乘288。"}, {"t0": 19.783, "t1": 20.97, "text": "为什么不要B帧?"}, {"t0": 20.97, "t1": 27.713, "text": "因为B帧要等未来帧,每多等一帧,对通话都是不可接受的。"}], "subs": [{"t0": 0.35, "t1": 2.35, "text": "1996年的H.263"}, {"t0": 2.37, "t1": 5.97, "text": "痛点是电话线和早期移动网络上的可视电话"}, {"t0": 5.99, "t1": 9.27, "text": "几十k的带宽,延迟就是一切"}, {"t0": 9.29, "t1": 13.53, "text": "它的基线档只有I帧和P帧,半像素运动矢量"}, {"t0": 13.55, "t1": 17.575, "text": "而且只支持几种固定尺寸,从128乘96"}, {"t0": 17.595, "t1": 19.763, "text": "到CIF的352乘288"}, {"t0": 19.783, "t1": 20.95, "text": "为什么不要B帧?"}, {"t0": 20.97, "t1": 24.59, "text": "因为B帧要等未来帧,每多等一帧"}, {"t0": 24.61, "t1": 27.713, "text": "对通话都是不可接受的"}]}, {"id": "s14", "ch": "H.263 · 1996", "scene": "rec", "codec": "h263", "rec": [[176, 192], [596, 604]], "text": "它的微调也全围绕低码率展开:H.263+用附件的方式开关工具,D附件是无限制运动矢量,F附件是重叠块运动补偿,J附件是去块滤波。码控几乎都用恒定码率,因为信道本身就是恒定的。看这段3G手机拍的片子:整条GOP只有I和P,重排曲线是一条零线。", "start": 318.714, "dur": 25.643, "voice": "voice/s14.wav", "vdur": 24.443375, "sents": [{"t0": 0.35, "t1": 13.17, "text": "它的微调也全围绕低码率展开:H.263+用附件的方式开关工具,D附件是无限制运动矢量,F附件是重叠块运动补偿,J附件是去块滤波。"}, {"t0": 13.17, "t1": 18.29, "text": "码控几乎都用恒定码率,因为信道本身就是恒定的。"}, {"t0": 18.29, "t1": 24.793, "text": "看这段3G手机拍的片子:整条GOP只有I和P,重排曲线是一条零线。"}], "subs": [{"t0": 0.35, "t1": 3.21, "text": "它的微调也全围绕低码率展开"}, {"t0": 3.23, "t1": 6.47, "text": "H.263+用附件的方式开关工具"}, {"t0": 6.49, "t1": 11.21, "text": "D附件是无限制运动矢量,F附件是重叠块运动补偿"}, {"t0": 11.23, "t1": 13.15, "text": "J附件是去块滤波"}, {"t0": 13.17, "t1": 18.27, "text": "码控几乎都用恒定码率,因为信道本身就是恒定的"}, {"t0": 18.29, "t1": 22.77, "text": "看这段3G手机拍的片子:整条GOP只有I和P"}, {"t0": 22.79, "t1": 24.793, "text": "重排曲线是一条零线"}]}, {"id": "s15", "ch": "MPEG-2 · 1995", "scene": "codec", "codec": "mpeg2", "rec": [[676, 704]], "sfx": ["ui/cartography_table/drawmap2"], "text": "同一时期的MPEG-2,为DVD和数字电视而生。它用档次乘级别的矩阵来划定能力:主档次主级别,正好覆盖DVD的720乘576;4:2:2档次留给演播室。基线选择的原理很朴素:解码芯片按档次级别来造,内容就按芯片能力来编。DVD的总线只有9.8兆,所以它用恒定码率,加IBBP的GOP结构,把复杂场景的压力交给量化器。", "start": 344.357, "dur": 31.763, "voice": "voice/s15.wav", "vdur": 30.563375, "sents": [{"t0": 0.35, "t1": 4.31, "text": "同一时期的MPEG-2,为DVD和数字电视而生。"}, {"t0": 4.31, "t1": 12.25, "text": "它用档次乘级别的矩阵来划定能力:主档次主级别,正好覆盖DVD的720乘576;"}, {"t0": 12.25, "t1": 14.79, "text": "4:2:2档次留给演播室。"}, {"t0": 14.79, "t1": 21.79, "text": "基线选择的原理很朴素:解码芯片按档次级别来造,内容就按芯片能力来编。"}, {"t0": 21.79, "t1": 30.913, "text": "DVD的总线只有9.8兆,所以它用恒定码率,加IBBP的GOP结构,把复杂场景的压力交给量化器。"}], "subs": [{"t0": 0.35, "t1": 4.29, "text": "同一时期的MPEG-2,为DVD和数字电视而生"}, {"t0": 4.31, "t1": 9.23, "text": "它用档次乘级别的矩阵来划定能力:主档次主级别"}, {"t0": 9.25, "t1": 12.23, "text": "正好覆盖DVD的720乘576"}, {"t0": 12.25, "t1": 14.77, "text": "4:2:2档次留给演播室"}, {"t0": 14.79, "t1": 19.67, "text": "基线选择的原理很朴素:解码芯片按档次级别来造"}, {"t0": 19.69, "t1": 21.77, "text": "内容就按芯片能力来编"}, {"t0": 21.79, "t1": 25.969, "text": "DVD的总线只有9.8兆,所以它用恒定码率"}, {"t0": 25.989, "t1": 28.11, "text": "加IBBP的GOP结构"}, {"t0": 28.13, "t1": 30.913, "text": "把复杂场景的压力交给量化器"}]}, {"id": "s16", "ch": "VP6 · 2003", "scene": "codec", "codec": "vp6", "rec": [[20, 60]], "sfx": ["ui/cartography_table/drawmap3"], "text": "2003年,On2的VP6成了Flash视频的心脏,撑起了早期的网络视频。它只有I帧和P帧,关键帧同时充当金帧,被后续帧长期参考。它的档次选择很实际:VP6-E是默认档;VP6-S降低解码复杂度,专门照顾高分辨率下跑不动的老电脑。", "start": 376.121, "dur": 23.293, "voice": "voice/s16.wav", "vdur": 22.093375, "sents": [{"t0": 0.35, "t1": 6.57, "text": "2003年,On2的VP6成了Flash视频的心脏,撑起了早期的网络视频。"}, {"t0": 6.57, "t1": 12.67, "text": "它只有I帧和P帧,关键帧同时充当金帧,被后续帧长期参考。"}, {"t0": 12.67, "t1": 16.57, "text": "它的档次选择很实际:VP6-E是默认档;"}, {"t0": 16.57, "t1": 22.443, "text": "VP6-S降低解码复杂度,专门照顾高分辨率下跑不动的老电脑。"}], "subs": [{"t0": 0.35, "t1": 1.45, "text": "2003年"}, {"t0": 1.47, "t1": 4.27, "text": "On2的VP6成了Flash视频的心脏"}, {"t0": 4.29, "t1": 6.55, "text": "撑起了早期的网络视频"}, {"t0": 6.57, "t1": 10.59, "text": "它只有I帧和P帧,关键帧同时充当金帧"}, {"t0": 10.61, "t1": 12.65, "text": "被后续帧长期参考"}, {"t0": 12.67, "t1": 16.55, "text": "它的档次选择很实际:VP6-E是默认档"}, {"t0": 16.57, "t1": 19.29, "text": "VP6-S降低解码复杂度"}, {"t0": 19.31, "t1": 22.443, "text": "专门照顾高分辨率下跑不动的老电脑"}]}, {"id": "s17", "ch": "VP6 · 2003", "scene": "rec", "codec": "vp6", "rec": [[110, 150]], "text": "VP6的微调逻辑,都写在当年转码服务的参数里:两遍变码率,最简单的段落至少拿到目标码率的百分之四十,最难的段落最多拿到四倍。当每像素比特数低于0.07,编码器会主动丢帧,用时间分辨率换画面质量。", "start": 399.414, "dur": 20.647, "voice": "voice/s17.wav", "vdur": 19.446688, "sents": [{"t0": 0.35, "t1": 12.67, "text": "VP6的微调逻辑,都写在当年转码服务的参数里:两遍变码率,最简单的段落至少拿到目标码率的百分之四十,最难的段落最多拿到四倍。"}, {"t0": 12.67, "t1": 19.797, "text": "当每像素比特数低于0.07,编码器会主动丢帧,用时间分辨率换画面质量。"}], "subs": [{"t0": 0.35, "t1": 4.81, "text": "VP6的微调逻辑,都写在当年转码服务的参数里"}, {"t0": 4.83, "t1": 6.45, "text": "两遍变码率"}, {"t0": 6.47, "t1": 10.131, "text": "最简单的段落至少拿到目标码率的百分之四十"}, {"t0": 10.151, "t1": 12.65, "text": "最难的段落最多拿到四倍"}, {"t0": 12.67, "t1": 17.41, "text": "当每像素比特数低于0.07,编码器会主动丢帧"}, {"t0": 17.43, "t1": 19.797, "text": "用时间分辨率换画面质量"}]}, {"id": "s18", "ch": "Theora · 2004", "scene": "codec", "codec": "theora", "rec": [[780, 812]], "sfx": ["ui/loom/take_result1"], "text": "2001年,On2把VP3.2捐给了Xiph基金会,几年后它被改写成完全开放、免专利费的Theora。因为继承了VP3的单一档位血统,Theora没有任何档次体系,调参只剩两个维度:0到10的质量档,或者一个目标码率。它没有标准化的环路滤波,低码率下,8乘8块的边界会直接残留在画面上。", "start": 420.061, "dur": 26.673, "voice": "voice/s18.wav", "vdur": 25.473375, "sents": [{"t0": 0.35, "t1": 9.07, "text": "2001年,On2把VP3.2捐给了Xiph基金会,几年后它被改写成完全开放、免专利费的Theora。"}, {"t0": 9.07, "t1": 19.41, "text": "因为继承了VP3的单一档位血统,Theora没有任何档次体系,调参只剩两个维度:0到10的质量档,或者一个目标码率。"}, {"t0": 19.41, "t1": 25.823, "text": "它没有标准化的环路滤波,低码率下,8乘8块的边界会直接残留在画面上。"}], "subs": [{"t0": 0.35, "t1": 1.51, "text": "2001年"}, {"t0": 1.53, "t1": 4.052, "text": "On2把VP3.2捐给了Xiph基金会"}, {"t0": 4.072, "t1": 6.545, "text": "几年后它被改写成完全开放"}, {"t0": 6.565, "t1": 9.05, "text": "免专利费的Theora"}, {"t0": 9.07, "t1": 11.89, "text": "因为继承了VP3的单一档位血统"}, {"t0": 11.91, "t1": 16.03, "text": "Theora没有任何档次体系,调参只剩两个维度"}, {"t0": 16.05, "t1": 19.39, "text": "0到10的质量档,或者一个目标码率"}, {"t0": 19.41, "t1": 22.63, "text": "它没有标准化的环路滤波,低码率下"}, {"t0": 22.65, "t1": 25.823, "text": "8乘8块的边界会直接残留在画面上"}]}, {"id": "s19", "ch": "H.264 / AVC · 2003", "scene": "codec", "codec": "avc", "rec": [[448, 480]], "sfx": ["block/beacon/activate"], "text": "真正的分水岭,是2003年的H.264。Wiegand等人的综述论文总结了它的革新:4乘4整数变换,避免了编解码失配;九种方向的帧内预测;四分之一像素运动补偿;多参考帧;环路去块滤波;以及Marpe等人提出的CABAC,基于上下文的自适应二进制算术编码,比传统变长编码再省一成左右。和MPEG-2相比,同画质码率大约减半。", "start": 446.734, "dur": 30.323, "voice": "voice/s19.wav", "vdur": 29.123375, "sents": [{"t0": 0.35, "t1": 3.85, "text": "真正的分水岭,是2003年的H.264。"}, {"t0": 3.85, "t1": 10.61, "text": "Wiegand等人的综述论文总结了它的革新:4乘4整数变换,避免了编解码失配;"}, {"t0": 10.61, "t1": 12.57, "text": "九种方向的帧内预测;"}, {"t0": 12.57, "t1": 14.63, "text": "四分之一像素运动补偿;"}, {"t0": 14.63, "t1": 15.85, "text": "多参考帧;"}, {"t0": 15.85, "t1": 17.31, "text": "环路去块滤波;"}, {"t0": 17.31, "t1": 25.85, "text": "以及Marpe等人提出的CABAC,基于上下文的自适应二进制算术编码,比传统变长编码再省一成左右。"}, {"t0": 25.85, "t1": 29.473, "text": "和MPEG-2相比,同画质码率大约减半。"}], "subs": [{"t0": 0.35, "t1": 3.83, "text": "真正的分水岭,是2003年的H.264"}, {"t0": 3.85, "t1": 7.19, "text": "Wiegand等人的综述论文总结了它的革新"}, {"t0": 7.21, "t1": 10.59, "text": "4乘4整数变换,避免了编解码失配"}, {"t0": 10.61, "t1": 12.55, "text": "九种方向的帧内预测"}, {"t0": 12.57, "t1": 14.61, "text": "四分之一像素运动补偿"}, {"t0": 14.63, "t1": 15.83, "text": "多参考帧"}, {"t0": 15.85, "t1": 17.29, "text": "环路去块滤波"}, {"t0": 17.31, "t1": 20.05, "text": "以及Marpe等人提出的CABAC"}, {"t0": 20.07, "t1": 23.05, "text": "基于上下文的自适应二进制算术编码"}, {"t0": 23.07, "t1": 25.83, "text": "比传统变长编码再省一成左右"}, {"t0": 25.85, "t1": 29.473, "text": "和MPEG-2相比,同画质码率大约减半"}]}, {"id": "s20", "ch": "H.264 / AVC · 2003", "scene": "profiles", "codec": "avc", "text": "H.264的档次选择,背后是复杂度和延迟的取舍。Baseline档砍掉了B帧和CABAC,解码简单、零重排延迟,所以至今还是视频通话的选择;Main档加入B帧和CABAC,是广播的主力;High档再加上8乘8变换和量化矩阵,成为点播和存档的默认。要10比特或者4:4:4,就继续往High10、High422、High444走。", "start": 477.057, "dur": 27.803, "voice": "voice/s20.wav", "vdur": 26.603375, "sents": [{"t0": 0.35, "t1": 4.75, "text": "H.264的档次选择,背后是复杂度和延迟的取舍。"}, {"t0": 4.75, "t1": 12.29, "text": "Baseline档砍掉了B帧和CABAC,解码简单、零重排延迟,所以至今还是视频通话的选择;"}, {"t0": 12.29, "t1": 16.31, "text": "Main档加入B帧和CABAC,是广播的主力;"}, {"t0": 16.31, "t1": 21.71, "text": "High档再加上8乘8变换和量化矩阵,成为点播和存档的默认。"}, {"t0": 21.71, "t1": 26.953, "text": "要10比特或者4:4:4,就继续往High10、High422、High444走。"}], "subs": [{"t0": 0.35, "t1": 4.73, "text": "H.264的档次选择,背后是复杂度和延迟的取舍"}, {"t0": 4.75, "t1": 7.59, "text": "Baseline档砍掉了B帧和CABAC"}, {"t0": 7.61, "t1": 9.61, "text": "解码简单、零重排延迟"}, {"t0": 9.63, "t1": 12.27, "text": "所以至今还是视频通话的选择"}, {"t0": 12.29, "t1": 16.29, "text": "Main档加入B帧和CABAC,是广播的主力"}, {"t0": 16.31, "t1": 19.49, "text": "High档再加上8乘8变换和量化矩阵"}, {"t0": 19.51, "t1": 21.69, "text": "成为点播和存档的默认"}, {"t0": 21.71, "t1": 24.81, "text": "要10比特或者4:4:4,就继续往High10"}, {"t0": 24.83, "t1": 26.953, "text": "High422、High444走"}]}, {"id": "s21", "ch": "H.264 / AVC · 2003", "scene": "tune", "codec": "avc", "rec": [[416, 448]], "text": "x264的微调是一套工程艺术。CRF恒定质量模式按内容复杂度分配比特,是存档首选;直播则用目标码率加VBV缓冲来限幅。preset决定搜索力度,从ultrafast到placebo,本质是在率失真优化里花多少算力。tune按片源微调:film保细节,animation加大B帧和去块,grain保颗粒,zerolatency关掉B帧和预读。还有Garrett-Glaser提出的宏块树算法:一个块被后续帧引用得越多,就越值得多给比特。", "start": 504.861, "dur": 35.88, "voice": "voice/s21.wav", "vdur": 34.680063, "sents": [{"t0": 0.35, "t1": 3.23, "text": "x264的微调是一套工程艺术。"}, {"t0": 3.23, "t1": 7.07, "text": "CRF恒定质量模式按内容复杂度分配比特,是存档首选;"}, {"t0": 7.07, "t1": 10.87, "text": "直播则用目标码率加VBV缓冲来限幅。"}, {"t0": 10.87, "t1": 19.01, "text": "preset决定搜索力度,从ultrafast到placebo,本质是在率失真优化里花多少算力。"}, {"t0": 19.01, "t1": 27.71, "text": "tune按片源微调:film保细节,animation加大B帧和去块,grain保颗粒,zerolatency关掉B帧和预读。"}, {"t0": 27.71, "t1": 35.03, "text": "还有Garrett-Glaser提出的宏块树算法:一个块被后续帧引用得越多,就越值得多给比特。"}], "subs": [{"t0": 0.35, "t1": 3.21, "text": "x264的微调是一套工程艺术"}, {"t0": 3.23, "t1": 7.05, "text": "CRF恒定质量模式按内容复杂度分配比特"}, {"t0": 7.07, "t1": 7.37, "text": "是存档首选"}, {"t0": 7.07, "t1": 10.85, "text": "直播则用目标码率加VBV缓冲来限幅"}, {"t0": 10.87, "t1": 12.15, "text": "preset决定搜索力度"}, {"t0": 12.17, "t1": 15.069, "text": "从ultrafast到placebo"}, {"t0": 15.089, "t1": 18.99, "text": "本质是在率失真优化里花多少算力"}, {"t0": 19.01, "t1": 20.89, "text": "tune按片源微调:film保细节"}, {"t0": 20.91, "t1": 23.85, "text": "animation加大B帧和去块"}, {"t0": 23.87, "t1": 25.23, "text": "grain保颗粒"}, {"t0": 25.25, "t1": 27.69, "text": "zerolatency关掉B帧和预读"}, {"t0": 27.71, "t1": 30.93, "text": "还有Garrett-Glaser提出的宏块树算法"}, {"t0": 30.95, "t1": 35.03, "text": "一个块被后续帧引用得越多,就越值得多给比特"}]}, {"id": "s22", "ch": "VP8 · 2010", "scene": "codec", "codec": "vp8", "rec": [[928, 960]], "sfx": ["block/amethyst/resonate2"], "text": "2010年谷歌收购On2,开源了VP8,配合WebM对抗H.264的专利费。Bankoski等人的技术综述里,最有特色的是替代参考帧AltRef:编码器偷看未来最多25帧,合成一个不显示、只用来被参考的帧,起到类似B帧的作用,却不打乱显示顺序。它的档次0到3只是解码复杂度的递减,WebM一律推荐0。直播用CBR,同时关掉AltRef和预读,换取零额外延迟。", "start": 540.741, "dur": 33.843, "voice": "voice/s22.wav", "vdur": 32.643375, "sents": [{"t0": 0.35, "t1": 7.29, "text": "2010年谷歌收购On2,开源了VP8,配合WebM对抗H.264的专利费。"}, {"t0": 7.29, "t1": 22.15, "text": "Bankoski等人的技术综述里,最有特色的是替代参考帧AltRef:编码器偷看未来最多25帧,合成一个不显示、只用来被参考的帧,起到类似B帧的作用,却不打乱显示顺序。"}, {"t0": 22.15, "t1": 27.73, "text": "它的档次0到3只是解码复杂度的递减,WebM一律推荐0。"}, {"t0": 27.73, "t1": 32.993, "text": "直播用CBR,同时关掉AltRef和预读,换取零额外延迟。"}], "subs": [{"t0": 0.35, "t1": 3.89, "text": "2010年谷歌收购On2,开源了VP8"}, {"t0": 3.91, "t1": 7.27, "text": "配合WebM对抗H.264的专利费"}, {"t0": 7.29, "t1": 9.51, "text": "Bankoski等人的技术综述里"}, {"t0": 9.53, "t1": 12.59, "text": "最有特色的是替代参考帧AltRef"}, {"t0": 12.61, "t1": 15.51, "text": "编码器偷看未来最多25帧,合成一个不显示"}, {"t0": 15.53, "t1": 20.35, "text": "只用来被参考的帧,起到类似B帧的作用"}, {"t0": 20.37, "t1": 22.13, "text": "却不打乱显示顺序"}, {"t0": 22.15, "t1": 25.49, "text": "它的档次0到3只是解码复杂度的递减"}, {"t0": 25.51, "t1": 27.71, "text": "WebM一律推荐0"}, {"t0": 27.73, "t1": 31.25, "text": "直播用CBR,同时关掉AltRef和预读"}, {"t0": 31.27, "t1": 32.993, "text": "换取零额外延迟"}]}, {"id": "s23", "ch": "H.265 / HEVC · 2013", "scene": "codec", "codec": "hevc", "rec": [[604, 632]], "sfx": ["block/beacon/power2"], "text": "2013年的H.265,也就是HEVC,目标是4K。Sullivan等人的综述描述了它的核心变化:宏块被64乘64的编码树单元取代,按四叉树递归切分;帧内预测方向从九种增加到三十五种;新增样点自适应偏移SAO。同画质下,比H.264再省大约百分之四十到五十。", "start": 574.584, "dur": 26.043, "voice": "voice/s23.wav", "vdur": 24.843375, "sents": [{"t0": 0.35, "t1": 5.29, "text": "2013年的H.265,也就是HEVC,目标是4K。"}, {"t0": 5.29, "t1": 14.049, "text": "Sullivan等人的综述描述了它的核心变化:宏块被64乘64的编码树单元取代,按四叉树递归切分;"}, {"t0": 14.049, "t1": 18.55, "text": "帧内预测方向从九种增加到三十五种;"}, {"t0": 18.55, "t1": 20.77, "text": "新增样点自适应偏移SAO。"}, {"t0": 20.77, "t1": 25.193, "text": "同画质下,比H.264再省大约百分之四十到五十。"}], "subs": [{"t0": 0.35, "t1": 3.77, "text": "2013年的H.265,也就是HEVC"}, {"t0": 3.79, "t1": 5.27, "text": "目标是4K"}, {"t0": 5.29, "t1": 8.73, "text": "Sullivan等人的综述描述了它的核心变化"}, {"t0": 8.75, "t1": 12.57, "text": "宏块被64乘64的编码树单元取代"}, {"t0": 12.59, "t1": 14.029, "text": "按四叉树递归切分"}, {"t0": 14.049, "t1": 18.53, "text": "帧内预测方向从九种增加到三十五种"}, {"t0": 18.55, "t1": 20.75, "text": "新增样点自适应偏移SAO"}, {"t0": 20.77, "t1": 25.193, "text": "同画质下,比H.264再省大约百分之四十到五十"}]}, {"id": "s24", "ch": "H.265 / HEVC · 2013", "scene": "rec", "codec": "hevc", "rec": [[632, 664]], "text": "HEVC的基线,一个关键选择是Main还是Main10。即使片源是8比特,用10比特内部精度编码,也能减少渐变天空里的色带,所以4K和HDR一律选Main10。x265的微调延续了x264的思路:CRF默认28,preset越慢压缩率越高;tune grain会开启颗粒保护和恒定VBV,让噪点不被抹平。", "start": 600.628, "dur": 27.203, "voice": "voice/s24.wav", "vdur": 26.003375, "sents": [{"t0": 0.35, "t1": 4.71, "text": "HEVC的基线,一个关键选择是Main还是Main10。"}, {"t0": 4.71, "t1": 13.89, "text": "即使片源是8比特,用10比特内部精度编码,也能减少渐变天空里的色带,所以4K和HDR一律选Main10。"}, {"t0": 13.89, "t1": 21.35, "text": "x265的微调延续了x264的思路:CRF默认28,preset越慢压缩率越高;"}, {"t0": 21.35, "t1": 26.353, "text": "tune grain会开启颗粒保护和恒定VBV,让噪点不被抹平。"}], "subs": [{"t0": 0.35, "t1": 1.73, "text": "HEVC的基线"}, {"t0": 1.75, "t1": 4.69, "text": "一个关键选择是Main还是Main10"}, {"t0": 4.71, "t1": 8.45, "text": "即使片源是8比特,用10比特内部精度编码"}, {"t0": 8.47, "t1": 10.83, "text": "也能减少渐变天空里的色带"}, {"t0": 10.85, "t1": 13.87, "text": "所以4K和HDR一律选Main10"}, {"t0": 13.89, "t1": 17.29, "text": "x265的微调延续了x264的思路"}, {"t0": 17.31, "t1": 21.33, "text": "CRF默认28,preset越慢压缩率越高"}, {"t0": 21.35, "t1": 24.57, "text": "tune grain会开启颗粒保护和恒定VBV"}, {"t0": 24.59, "t1": 26.353, "text": "让噪点不被抹平"}]}, {"id": "s25", "ch": "VP9 · 2013", "scene": "codec", "codec": "vp9", "rec": [[960, 984]], "sfx": ["block/amethyst/resonate3"], "text": "同一年谷歌发布VP9,被YouTube大规模采用。它引入64乘64超级块和最多8个分段。档次按采样和位深划分:档次0是8比特4:2:0,档次2是10比特HDR,1和3留给4:4:4。微调上要注意tile:把画面切成多列可以并行编码,但列边界会阻断预测,码率会上涨百分之一到三;而行级多线程没有码率代价,应该常开。", "start": 627.831, "dur": 30.683, "voice": "voice/s25.wav", "vdur": 29.483375, "sents": [{"t0": 0.35, "t1": 4.29, "text": "同一年谷歌发布VP9,被YouTube大规模采用。"}, {"t0": 4.29, "t1": 8.55, "text": "它引入64乘64超级块和最多8个分段。"}, {"t0": 8.55, "t1": 17.19, "text": "档次按采样和位深划分:档次0是8比特4:2:0,档次2是10比特HDR,1和3留给4:4:4。"}, {"t0": 17.19, "t1": 26.21, "text": "微调上要注意tile:把画面切成多列可以并行编码,但列边界会阻断预测,码率会上涨百分之一到三;"}, {"t0": 26.21, "t1": 29.833, "text": "而行级多线程没有码率代价,应该常开。"}], "subs": [{"t0": 0.35, "t1": 2.3, "text": "同一年谷歌发布VP9"}, {"t0": 2.32, "t1": 4.27, "text": "被YouTube大规模采用"}, {"t0": 4.29, "t1": 8.53, "text": "它引入64乘64超级块和最多8个分段"}, {"t0": 8.55, "t1": 13.05, "text": "档次按采样和位深划分:档次0是8比特4:2:0"}, {"t0": 13.07, "t1": 17.17, "text": "档次2是10比特HDR,1和3留给4:4:4"}, {"t0": 17.19, "t1": 18.91, "text": "微调上要注意tile"}, {"t0": 18.93, "t1": 23.89, "text": "把画面切成多列可以并行编码,但列边界会阻断预测"}, {"t0": 23.91, "t1": 26.19, "text": "码率会上涨百分之一到三"}, {"t0": 26.21, "t1": 29.833, "text": "而行级多线程没有码率代价,应该常开"}]}, {"id": "s26", "ch": "AV1 · 2018", "scene": "codec", "codec": "av1", "rec": [[224, 260]], "sfx": ["block/beacon/power3"], "text": "2018年,开放媒体联盟发布AV1,整合了谷歌VP10、思科Thor和Mozilla Daala的成果。它的超级块达到128乘128,帧内方向细化到56种,还有从亮度预测色度的CfL、约束方向增强滤波CDEF,以及环路恢复滤波。AV1没有传统意义的B帧,分析工具里全部显示为P帧,但它用隐藏的替代参考帧,实现了同样的双向预测。", "start": 658.514, "dur": 30.193, "voice": "voice/s26.wav", "vdur": 28.993375, "sents": [{"t0": 0.35, "t1": 8.01, "text": "2018年,开放媒体联盟发布AV1,整合了谷歌VP10、思科Thor和Mozilla Daala的成果。"}, {"t0": 8.01, "t1": 20.09, "text": "它的超级块达到128乘128,帧内方向细化到56种,还有从亮度预测色度的CfL、约束方向增强滤波CDEF,以及环路恢复滤波。"}, {"t0": 20.09, "t1": 29.343, "text": "AV1没有传统意义的B帧,分析工具里全部显示为P帧,但它用隐藏的替代参考帧,实现了同样的双向预测。"}], "subs": [{"t0": 0.35, "t1": 3.61, "text": "2018年,开放媒体联盟发布AV1"}, {"t0": 3.63, "t1": 5.149, "text": "整合了谷歌VP10"}, {"t0": 5.169, "t1": 7.99, "text": "思科Thor和Mozilla Daala的成果"}, {"t0": 8.01, "t1": 11.25, "text": "它的超级块达到128乘128"}, {"t0": 11.27, "t1": 13.67, "text": "帧内方向细化到56种"}, {"t0": 13.69, "t1": 15.446, "text": "还有从亮度预测色度的CfL"}, {"t0": 15.466, "t1": 20.07, "text": "约束方向增强滤波CDEF,以及环路恢复滤波"}, {"t0": 20.09, "t1": 22.53, "text": "AV1没有传统意义的B帧"}, {"t0": 22.55, "t1": 25.07, "text": "分析工具里全部显示为P帧"}, {"t0": 25.09, "t1": 29.343, "text": "但它用隐藏的替代参考帧,实现了同样的双向预测"}]}, {"id": "s27", "ch": "AV1 · 2018", "scene": "grain", "codec": "av1", "rec": [[300, 340]], "text": "最精彩的是胶片颗粒合成。Norkin和Birkbeck在2018年的论文里提出:先把颗粒去掉,只编码干净的画面,再在帧头写一张不到一百字节的颗粒参数表,让解码器自己把颗粒画回去。颗粒是最难压缩的随机噪声,这一招省下了大量码率。SVT-AV1的preset从0到13,数字越大越快;tune选0追求主观视觉,选1追求PSNR指标。同等VMAF下,AV1比x265大约再省百分之三十。", "start": 688.708, "dur": 37.443, "voice": "voice/s27.wav", "vdur": 36.243375, "sents": [{"t0": 0.35, "t1": 2.73, "text": "最精彩的是胶片颗粒合成。"}, {"t0": 2.73, "t1": 17.665, "text": "Norkin和Birkbeck在2018年的论文里提出:先把颗粒去掉,只编码干净的画面,再在帧头写一张不到一百字节的颗粒参数表,让解码器自己把颗粒画回去。"}, {"t0": 17.665, "t1": 22.953, "text": "颗粒是最难压缩的随机噪声,这一招省下了大量码率。"}, {"t0": 22.953, "t1": 27.439, "text": "SVT-AV1的preset从0到13,数字越大越快;"}, {"t0": 27.439, "t1": 31.55, "text": "tune选0追求主观视觉,选1追求PSNR指标。"}, {"t0": 31.55, "t1": 36.593, "text": "同等VMAF下,AV1比x265大约再省百分之三十。"}], "subs": [{"t0": 0.35, "t1": 2.71, "text": "最精彩的是胶片颗粒合成"}, {"t0": 2.73, "t1": 6.67, "text": "Norkin和Birkbeck在2018年的论文里提出"}, {"t0": 6.69, "t1": 10.17, "text": "先把颗粒去掉,只编码干净的画面"}, {"t0": 10.19, "t1": 14.686, "text": "再在帧头写一张不到一百字节的颗粒参数表"}, {"t0": 14.706, "t1": 17.645, "text": "让解码器自己把颗粒画回去"}, {"t0": 17.665, "t1": 22.933, "text": "颗粒是最难压缩的随机噪声,这一招省下了大量码率"}, {"t0": 22.953, "t1": 24.75, "text": "SVT-AV1的preset从0到13"}, {"t0": 24.77, "t1": 27.419, "text": "数字越大越快"}, {"t0": 27.439, "t1": 31.53, "text": "tune选0追求主观视觉,选1追求PSNR指标"}, {"t0": 31.55, "t1": 33.15, "text": "同等VMAF下"}, {"t0": 33.17, "t1": 36.593, "text": "AV1比x265大约再省百分之三十"}]}, {"id": "s28", "ch": "H.266 / VVC · 2020", "scene": "codec", "codec": "vvc", "rec": [[1000, 1012], [0, 8]], "sfx": ["block/enchantment_table/enchant2"], "text": "2020年定稿的H.266,全名通用视频编码VVC。Bross等人的综述给出数字:比HEVC省约百分之五十,比H.264省约百分之七十五。它的编码树单元扩大到128,在四叉树之外加入二叉和三叉的多类型树;帧内方向增加到67种;还有仿射运动补偿、几何划分、自适应环路滤波ALF,以及为屏幕内容设计的帧内块拷贝IBC。录屏开头那段腾讯的IBC测试流,就是它。", "start": 726.151, "dur": 34.323, "voice": "voice/s28.wav", "vdur": 33.123375, "sents": [{"t0": 0.35, "t1": 5.09, "text": "2020年定稿的H.266,全名通用视频编码VVC。"}, {"t0": 5.09, "t1": 12.77, "text": "Bross等人的综述给出数字:比HEVC省约百分之五十,比H.264省约百分之七十五。"}, {"t0": 12.77, "t1": 18.99, "text": "它的编码树单元扩大到128,在四叉树之外加入二叉和三叉的多类型树;"}, {"t0": 18.99, "t1": 21.23, "text": "帧内方向增加到67种;"}, {"t0": 21.23, "t1": 29.55, "text": "还有仿射运动补偿、几何划分、自适应环路滤波ALF,以及为屏幕内容设计的帧内块拷贝IBC。"}, {"t0": 29.55, "t1": 33.473, "text": "录屏开头那段腾讯的IBC测试流,就是它。"}], "subs": [{"t0": 0.35, "t1": 2.69, "text": "2020年定稿的H.266"}, {"t0": 2.71, "t1": 5.07, "text": "全名通用视频编码VVC"}, {"t0": 5.09, "t1": 7.41, "text": "Bross等人的综述给出数字"}, {"t0": 7.43, "t1": 9.87, "text": "比HEVC省约百分之五十"}, {"t0": 9.89, "t1": 12.75, "text": "比H.264省约百分之七十五"}, {"t0": 12.77, "t1": 15.37, "text": "它的编码树单元扩大到128"}, {"t0": 15.39, "t1": 18.97, "text": "在四叉树之外加入二叉和三叉的多类型树"}, {"t0": 18.99, "t1": 21.21, "text": "帧内方向增加到67种"}, {"t0": 21.23, "t1": 23.828, "text": "还有仿射运动补偿、几何划分"}, {"t0": 23.848, "t1": 25.39, "text": "自适应环路滤波ALF"}, {"t0": 25.41, "t1": 29.53, "text": "以及为屏幕内容设计的帧内块拷贝IBC"}, {"t0": 29.55, "t1": 33.473, "text": "录屏开头那段腾讯的IBC测试流,就是它"}]}, {"id": "s29", "ch": "H.266 / VVC · 2020", "scene": "mtt", "codec": "vvc", "rec": [[984, 1000]], "text": "VVC的基线一开始就是Main10,10比特成为起点,而不是选项;需要4:4:4就选Main10 4:4:4,需要可分级就选多层档次。开源编码器VVenC把微调收敛成五个档位,从faster到slower,构成一组帕累托最优;qpa参数决定是按人眼感知分配量化,还是按PSNR追指标。", "start": 760.474, "dur": 25.823, "voice": "voice/s29.wav", "vdur": 24.623375, "sents": [{"t0": 0.35, "t1": 6.51, "text": "VVC的基线一开始就是Main10,10比特成为起点,而不是选项;"}, {"t0": 6.51, "t1": 12.17, "text": "需要4:4:4就选Main10 4:4:4,需要可分级就选多层档次。"}, {"t0": 12.17, "t1": 19.55, "text": "开源编码器VVenC把微调收敛成五个档位,从faster到slower,构成一组帕累托最优;"}, {"t0": 19.55, "t1": 24.973, "text": "qpa参数决定是按人眼感知分配量化,还是按PSNR追指标。"}], "subs": [{"t0": 0.35, "t1": 3.29, "text": "VVC的基线一开始就是Main10"}, {"t0": 3.31, "t1": 6.49, "text": "10比特成为起点,而不是选项"}, {"t0": 6.51, "t1": 9.25, "text": "需要4:4:4就选Main10 4:4:4"}, {"t0": 9.27, "t1": 12.15, "text": "需要可分级就选多层档次"}, {"t0": 12.17, "t1": 16.11, "text": "开源编码器VVenC把微调收敛成五个档位"}, {"t0": 16.13, "t1": 17.61, "text": "从faster到slower"}, {"t0": 17.63, "t1": 19.53, "text": "构成一组帕累托最优"}, {"t0": 19.55, "t1": 22.79, "text": "qpa参数决定是按人眼感知分配量化"}, {"t0": 22.81, "t1": 24.973, "text": "还是按PSNR追指标"}]}, {"id": "s30", "ch": "AVS3 · 2019", "scene": "codec", "codec": "avs3", "rec": [[500, 540]], "sfx": ["block/bell/bell_use01"], "text": "中国的AVS3,第一阶段在2019年定稿。北京大学团队的综述显示,它在4K序列上比AVS2省约百分之二十四,比HEVC省约百分之二十七。划分上,它在四叉树、二叉树之外引入了扩展四叉树EQT,还有极致运动矢量表达UMVE、角度加权预测AWP等工具,并在2022年被DVB采纳为下一代超高清标准之一。", "start": 786.298, "dur": 29.303, "voice": "voice/s30.wav", "vdur": 28.103375, "sents": [{"t0": 0.35, "t1": 4.41, "text": "中国的AVS3,第一阶段在2019年定稿。"}, {"t0": 4.41, "t1": 13.39, "text": "北京大学团队的综述显示,它在4K序列上比AVS2省约百分之二十四,比HEVC省约百分之二十七。"}, {"t0": 13.39, "t1": 28.453, "text": "划分上,它在四叉树、二叉树之外引入了扩展四叉树EQT,还有极致运动矢量表达UMVE、角度加权预测AWP等工具,并在2022年被DVB采纳为下一代超高清标准之一。"}], "subs": [{"t0": 0.35, "t1": 4.39, "text": "中国的AVS3,第一阶段在2019年定稿"}, {"t0": 4.41, "t1": 6.79, "text": "北京大学团队的综述显示"}, {"t0": 6.81, "t1": 10.454, "text": "它在4K序列上比AVS2省约百分之二十四"}, {"t0": 10.474, "t1": 13.37, "text": "比HEVC省约百分之二十七"}, {"t0": 13.39, "t1": 15.553, "text": "划分上,它在四叉树"}, {"t0": 15.573, "t1": 18.39, "text": "二叉树之外引入了扩展四叉树EQT"}, {"t0": 18.41, "t1": 21.01, "text": "还有极致运动矢量表达UMVE"}, {"t0": 21.03, "t1": 23.77, "text": "角度加权预测AWP等工具"}, {"t0": 23.79, "t1": 28.453, "text": "并在2022年被DVB采纳为下一代超高清标准之一"}]}, {"id": "s31", "ch": "AVS3 · 2019", "scene": "rec", "codec": "avs3", "rec": [[548, 576]], "text": "AVS3的基线:Main是8比特,Main10是4K广播的主力,High在第二阶段加入更多工具。微调上,参考软件提供两套配置:随机访问配置用分层B帧和32帧的大GOP,效率优先,适合点播和广播;低延迟配置只做前向参考,适合直播,代价是同画质码率高出一到两成。", "start": 815.601, "dur": 26.363, "voice": "voice/s31.wav", "vdur": 25.163375, "sents": [{"t0": 0.35, "t1": 8.43, "text": "AVS3的基线:Main是8比特,Main10是4K广播的主力,High在第二阶段加入更多工具。"}, {"t0": 8.43, "t1": 19.15, "text": "微调上,参考软件提供两套配置:随机访问配置用分层B帧和32帧的大GOP,效率优先,适合点播和广播;"}, {"t0": 19.15, "t1": 25.513, "text": "低延迟配置只做前向参考,适合直播,代价是同画质码率高出一到两成。"}], "subs": [{"t0": 0.35, "t1": 3.17, "text": "AVS3的基线:Main是8比特"}, {"t0": 3.19, "t1": 5.59, "text": "Main10是4K广播的主力"}, {"t0": 5.61, "t1": 8.41, "text": "High在第二阶段加入更多工具"}, {"t0": 8.43, "t1": 11.85, "text": "微调上,参考软件提供两套配置"}, {"t0": 11.87, "t1": 15.85, "text": "随机访问配置用分层B帧和32帧的大GOP"}, {"t0": 15.87, "t1": 19.13, "text": "效率优先,适合点播和广播"}, {"t0": 19.15, "t1": 22.61, "text": "低延迟配置只做前向参考,适合直播"}, {"t0": 22.63, "t1": 25.513, "text": "代价是同画质码率高出一到两成"}]}, {"id": "s32", "ch": "AV2 · 2026", "scene": "av2", "codec": "av2", "sfx": ["fireworks/largeblast1", "fireworks/twinkle_far1"], "text": "最后是AV2。开放媒体联盟在2026年发布了1.0规范。Norkin的综述报告:随机访问配置下,它比AV1在PSNR上省约百分之二十九,在VMAF上省约百分之三十三。超级块扩大到256,新工具包括二次变换、网格编码量化、奇偶隐藏和跨色度分量变换。不过目前只有参考编码器,只支持单遍,速度远低于每秒一帧,档次也只定义到10比特。", "start": 841.965, "dur": 32.293, "voice": "voice/s32.wav", "vdur": 31.093375, "sents": [{"t0": 0.35, "t1": 1.75, "text": "最后是AV2。"}, {"t0": 1.75, "t1": 5.505, "text": "开放媒体联盟在2026年发布了1.0规范。"}, {"t0": 5.505, "t1": 15.173, "text": "Norkin的综述报告:随机访问配置下,它比AV1在PSNR上省约百分之二十九,在VMAF上省约百分之三十三。"}, {"t0": 15.173, "t1": 23.65, "text": "超级块扩大到256,新工具包括二次变换、网格编码量化、奇偶隐藏和跨色度分量变换。"}, {"t0": 23.65, "t1": 31.443, "text": "不过目前只有参考编码器,只支持单遍,速度远低于每秒一帧,档次也只定义到10比特。"}], "subs": [{"t0": 0.35, "t1": 1.73, "text": "最后是AV2"}, {"t0": 1.75, "t1": 5.485, "text": "开放媒体联盟在2026年发布了1.0规范"}, {"t0": 5.505, "t1": 9.31, "text": "Norkin的综述报告:随机访问配置下"}, {"t0": 9.33, "t1": 12.313, "text": "它比AV1在PSNR上省约百分之二十九"}, {"t0": 12.333, "t1": 15.153, "text": "在VMAF上省约百分之三十三"}, {"t0": 15.173, "t1": 18.17, "text": "超级块扩大到256,新工具包括二次变换"}, {"t0": 18.19, "t1": 23.63, "text": "网格编码量化、奇偶隐藏和跨色度分量变换"}, {"t0": 23.65, "t1": 27.35, "text": "不过目前只有参考编码器,只支持单遍"}, {"t0": 27.37, "t1": 31.443, "text": "速度远低于每秒一帧,档次也只定义到10比特"}]}, {"id": "s33", "ch": "四 · 总结", "scene": "evolution", "sfx": ["block/amethyst/shimmer"], "text": "回头看这三十年,技术迭代有几条清晰的主线。块越来越大:16、64、128,再到256。划分越来越灵活:从固定宏块,到四叉树,再到多类型树。预测越来越细:帧内方向从9到35再到67,运动矢量从半像素走到十六分之一像素,还有仿射运动。滤波越来越多:从没有,到去块,再到SAO、ALF、CDEF。熵编码则从变长码,走向了自适应算术编码。", "start": 874.258, "dur": 34.813, "voice": "voice/s33.wav", "vdur": 33.613375, "sents": [{"t0": 0.35, "t1": 4.69, "text": "回头看这三十年,技术迭代有几条清晰的主线。"}, {"t0": 4.69, "t1": 9.45, "text": "块越来越大:16、64、128,再到256。"}, {"t0": 9.45, "t1": 14.93, "text": "划分越来越灵活:从固定宏块,到四叉树,再到多类型树。"}, {"t0": 14.93, "t1": 23.13, "text": "预测越来越细:帧内方向从9到35再到67,运动矢量从半像素走到十六分之一像素,还有仿射运动。"}, {"t0": 23.13, "t1": 29.81, "text": "滤波越来越多:从没有,到去块,再到SAO、ALF、CDEF。"}, {"t0": 29.81, "t1": 33.963, "text": "熵编码则从变长码,走向了自适应算术编码。"}], "subs": [{"t0": 0.35, "t1": 4.67, "text": "回头看这三十年,技术迭代有几条清晰的主线"}, {"t0": 4.69, "t1": 9.43, "text": "块越来越大:16、64、128,再到256"}, {"t0": 9.45, "t1": 13.41, "text": "划分越来越灵活:从固定宏块,到四叉树"}, {"t0": 13.43, "t1": 14.91, "text": "再到多类型树"}, {"t0": 14.93, "t1": 18.589, "text": "预测越来越细:帧内方向从9到35再到67"}, {"t0": 18.609, "t1": 22.001, "text": "运动矢量从半像素走到十六分之一像素"}, {"t0": 22.021, "t1": 23.11, "text": "还有仿射运动"}, {"t0": 23.13, "t1": 27.93, "text": "滤波越来越多:从没有,到去块,再到SAO"}, {"t0": 27.95, "t1": 29.79, "text": "ALF、CDEF"}, {"t0": 29.81, "t1": 33.963, "text": "熵编码则从变长码,走向了自适应算术编码"}]}, {"id": "s34", "ch": "四 · 总结", "scene": "choose", "sfx": ["random/orb"], "text": "每一代大约省一半码率,代价是编码复杂度翻上十倍。所以选格式不是越新越好:要兼容一切,选H.264;要效率又要硬件解码,选HEVC或AV1;为未来做准备,再看H.266、AVS3和AV2。", "start": 909.071, "dur": 19.357, "voice": "voice/s34.wav", "vdur": 18.156688, "sents": [{"t0": 0.35, "t1": 4.93, "text": "每一代大约省一半码率,代价是编码复杂度翻上十倍。"}, {"t0": 4.93, "t1": 9.97, "text": "所以选格式不是越新越好:要兼容一切,选H.264;"}, {"t0": 9.97, "t1": 13.91, "text": "要效率又要硬件解码,选HEVC或AV1;"}, {"t0": 13.91, "t1": 18.507, "text": "为未来做准备,再看H.266、AVS3和AV2。"}], "subs": [{"t0": 0.35, "t1": 4.91, "text": "每一代大约省一半码率,代价是编码复杂度翻上十倍"}, {"t0": 4.93, "t1": 8.55, "text": "所以选格式不是越新越好:要兼容一切"}, {"t0": 8.57, "t1": 9.95, "text": "选H.264"}, {"t0": 9.97, "t1": 13.89, "text": "要效率又要硬件解码,选HEVC或AV1"}, {"t0": 13.91, "t1": 18.507, "text": "为未来做准备,再看H.266、AVS3和AV2"}]}, {"id": "s35", "ch": "结尾", "scene": "outro", "rec": [[880, 900]], "sfx": ["ui/toast/challenge_complete"], "text": "压缩的本质,是在比特、画质和算力之间做交易。理解了I、P、B,理解了率失真,你就能看懂每一个编码参数背后的那笔账。本期用到的分析工具VideoAnalyzer-Qt是开源的,欢迎拿你自己的视频去拆一拆。我们下期见。", "start": 928.428, "dur": 19.307, "voice": "voice/s35.wav", "vdur": 18.106688, "sents": [{"t0": 0.35, "t1": 4.25, "text": "压缩的本质,是在比特、画质和算力之间做交易。"}, {"t0": 4.25, "t1": 10.51, "text": "理解了I、P、B,理解了率失真,你就能看懂每一个编码参数背后的那笔账。"}, {"t0": 10.51, "t1": 17.19, "text": "本期用到的分析工具VideoAnalyzer-Qt是开源的,欢迎拿你自己的视频去拆一拆。"}, {"t0": 17.19, "t1": 18.457, "text": "我们下期见。"}], "subs": [{"t0": 0.35, "t1": 4.23, "text": "压缩的本质,是在比特、画质和算力之间做交易"}, {"t0": 4.25, "t1": 7.05, "text": "理解了I、P、B,理解了率失真"}, {"t0": 7.07, "t1": 10.49, "text": "你就能看懂每一个编码参数背后的那笔账"}, {"t0": 10.51, "t1": 14.77, "text": "本期用到的分析工具VideoAnalyzer-Qt是开源的"}, {"t0": 14.79, "t1": 17.17, "text": "欢迎拿你自己的视频去拆一拆"}, {"t0": 17.19, "t1": 18.457, "text": "我们下期见"}]}, {"id": "s36", "ch": "参考文献", "scene": "refs", "text": "", "start": 947.735, "dur": 9.0, "sents": [], "subs": [], "sfx": ["item/book/close_put1"]}]};