-
Notifications
You must be signed in to change notification settings - Fork 0
/
README
211 lines (162 loc) · 9.31 KB
/
README
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
<!-- Author's note; this was intended to become a home page at some point,
but I'm not interested enough in grokking HTML right now, so it became
the README instead. Most of it should be valid Markdown. -->
Announcing Movit
================
Movit is the Modern Video Toolkit, notwithstanding that anything that's
called “modern” usually isn't, and it's really not a toolkit.
Movit aims to be a _high-quality_, _high-performance_, _open-source_
library for video filters.
TL;DR, please give me download link and system demands
======================================================
OK, you need
* A C++98 compiler. GCC will do. (I haven't tried Windows, but it
works fine on Linux and OS X, and Movit is not very POSIX-bound.)
* GNU Make.
* A GPU capable of running GLSL fragment shaders,
processing floating-point textures, and a few other things (all are
part of OpenGL 3.0 or newer, although most OpenGL 2.0 cards also
have what's needed through extensions). If your machine is less than five
years old _and you have the appropriate drivers_, you're home free.
GLES3 (for mobile devices) will also work.
* The [Eigen 3], [FFTW3] and [Google Test] libraries. (The library itself
does not depend on the latter, but you probably want to run the unit tests.)
* The [epoxy] library, for dealing with OpenGL extensions on various
platforms.
Movit has been tested with Intel GPUs with the Mesa drivers
(you'll probably need at least Mesa 8.0), Radeon 3850 and GeForce GTX 550
on Linux with the manufacturer's drivers, and with GeForce 8800 on OS X.
Again, most likely, GPU compatibility shouldn't be a big issue. See below
for performance estimates.
Still TL;DR, please give me the list of filters
===============================================
Blur, diffusion, FFT-based convolution, glow, lift/gamma/gain (color
correction), mirror, mix (add two inputs), luma mix (use a map to wipe between
two inputs), overlay (the Porter-Duff “over” operation), scale (bilinear and
Lanczos), sharpen (both by unsharp mask and by Wiener filters), saturation
(or desaturation), vignette, white balance, and a deinterlacer (YADIF).
Yes, that's a short list. But they all look great, are fast and don't give
you any nasty surprises. (I'd love to include denoise and
framerate up-/downconversion to the list, but doing them well are
all research-grade problems, and Movit is currently not there.)
TL;DR, but I am interested in a programming example instead
===========================================================
Assuming you have an OpenGL context already set up (either a classic OpenGL
context, a GL 3.x forward-compatible or core context, or a GLES3 context):
<code>
using namespace movit;
EffectChain chain(1280, 720);
ImageFormat inout_format;
inout_format.color_space = COLORSPACE_sRGB;
inout_format.gamma_curve = GAMMA_sRGB;
FlatInput *input = new FlatInput(inout_format, FORMAT_BGRA_POSTMULTIPLIED_ALPHA, GL_UNSIGNED_BYTE, 1280, 720));
chain.add_input(input);
Effect *saturation_effect = chain.add_effect(new SaturationEffect());
saturation_effect->set_float("saturation", 0.7f);
Effect *lift_gamma_gain_effect = chain.add_effect(new LiftGammaGainEffect());
const float gain[] = { 0.8f, 1.0f, 1.0f };
lift_gamma_gain_effect->set_vec3("gain", &gain);
chain.add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
chain.finalize();
for ( ;; ) {
// Do whatever you need here to decode the next frame into <pixels>.
input->set_pixel_data(pixels);
chain.render_to_screen();
}
</code>
OK, I can read a bit. What do you mean by “modern”?
===================================================
Backwards compatibility is fine and all, but sometimes we can do better
by observing that the world has moved on. In particular:
* It's 2016, so people want to edit HD video.
* It's 2016, so everybody has a GPU.
* It's 2016, so everybody has a working C++ compiler.
(Even Microsoft fixed theirs around 2003!)
While from a programming standpoint I'd love to say that it's 2016
and interlacing does no longer exist, but that's not true (and interlacing,
hated as it might be, is actually a useful and underrated technique for
bandwidth reduction in broadcast video). Movit may eventually provide
limited support for working with interlaced video; it has a deinterlacer,
but cannot currently process video in interlaced form.
What do you mean by “high-performance”?
=======================================
Today, you can hardly get a _cellphone_ without a multi-core, SIMD-capable
CPU, and a GPU. Yet, almost all open-source pixel processing I've seen
is written using straight-up single-threaded, scalar C! Clearly there is
room for improvement here, and that improvement is sorely needed.
We want to edit 1080p video, not watch slideshows.
Movit has chosen to run all pixel processing on the GPU, using GLSL—OpenCL is
way too young, and CUDA is single-vendor (and also surprisingly hard to
get good performance from for anything nontrivial). While “run on the GPU”
does not equal “infinite speed” (I am fairly certain that for many common
filters, I can beat the Intel-based GPU in my laptop with multithreaded SSE
code on the CPU—especially as moving the data to and from the GPU has a cost that is not
to be taken lightly), GPU programming is probably the _simplest_ way of writing
highly parallel code, and it also frees the CPU to do other things like video
decoding.
Exactly what speeds you can expect is of course highly dependent on
your GPU and the exact filter chain you are running. As a rule of thumb,
you can run a reasonable filter chain (a lift/gamma/gain operation,
a bit of diffusion, maybe a vignette) at 720p in around 30 fps on a four-year-old
Intel laptop. If you have a somewhat newer Intel card, you can do 1080p
video without much problems. And on a low-range nVidia card of today
(GTX 550 Ti), you can probably process 4K movies directly.
What do you mean by “high-quality”?
===================================
Movit aims to be high-quality in two important aspects, namely _code quality_
and _output quality_. (Unfortunately, documentation quality is not on the
list yet. Sorry.)
High-quality output?
====================
Movit works internally in linear floating-point all the way, strongly
reducing interim round-off and clipping errors. Furthermore, Movit is
(weakly) colorspace-aware. Why do colorspaces matter? Well, here's a video frame from a typical
camera, which records in Rec. 709 (the typical HDTV color space), and here's the
same frame misinterpreted as Rec. 601 (the typical SDTV color space):
[insert picture here]
The difference might be subtle, but would you like that color cast?
Maybe you could correct for it manually, but what if it happened on output
instead of on input? And I can promise you that once we move to more
wide-gamut color spaces, like the one in Rec. 2020 (used for UHDTV), the
difference will be anything but subtle. As of [why working in linear
light matters](http://www.4p8.com/eric.brasseur/gamma.html),
others have explained it better than I can; note also that this makes Movit
future-proof when the world moves towards 10- and 12-bit color precision
(although the latter requires Movit to change from 16-bit to 32-bit floating
point, it is a simple switch). The extra power from the GPU makes all of this
simple, so do we not need to make too many concessions for the sake of speed.
Movit does not currently do ICC profiles or advanced gamut mapping;
if you have out-of-gamut colors, they will clip. Sorry.
OK, and high-quality code?
==========================
Image processing code can be surprisingly subtle; it's easy to write
code that looks right, but that makes subtle artifacts that explode
when processed further in a later step. (Or code that simply never
worked, just that nobody cared to look at the output when a given
parameter was set. I've seen that, too.)
Movit tries to counteract this by three different strategies:
* First, _look at the output_. Does it look good? Really?
Even if you zoom in on the results? Don't settle for “meh, I'm
sure that's the best it can get”.
* Second, _keep things simple_. Movit does not aim for including
every possible video effect under the sun (there are [others out there]
that want that); the [YAGNI] principle is applied quite strongly throughout
the code. It's much better to write less code but actually
understand what it does; whenever I can replace some magic matrix
or obscure formula from the web with a clean calculation and a descriptive
comment on top, it makes me a bit happier. (Most of the time,
it turns out that I had used the matrix or formula in a wrong
way anyway. My degree is in multimedia signal processing, but it
does not mean I have a deep understanding of everything people do
in graphics.)
* Third, _have unit tests_. Tests are boring, but they are unforgiving
(much more unforgiving than your eye), and they keep stuff from breaking
afterwards. Almost every single test I wrote has uncovered bugs in Movit,
so they have already paid for themselves.
There is, of course, always room for improvement. I'm sure you can find
things that are stupid, little-thought-out, or buggy. If so, please let me
know.
What do you mean by “open-source”?
==================================
Movit is licensed under the [GNU GPL](http://www.gnu.org/licenses/gpl.html),
either version 2 or (at your option) any later version.