loopbio blog

New Open Source Releases: Imgstore and Stitchup

Written on Monday February 25, 2019

Over the last couple of months we released a couple of interesting open source tools that might be of use to some hackers out there. In typical bad-at-marketing-busy-company form we then forgot to blog about them!

Stitchup is a python library for stitching multiple videos together into a larger video mosaic or video panorama. This is a common scenario for users of Motif video recording systems who have multiple-camera setups to increase their spatial and temporal resolution.


Stitchup adds some functionality (storing and loading of calibrations) and considerably improves upon the OpenCV in-tree stitcher module Python API. Stitchup was used in the example above and for our previous post on multi-camera high throughput c.elegans screening assay

Imgstore is a seekable and scalable for video frames and metadata (see announcement and imgstore introduction here). Well, we just released version 2.0 (and then 2.1, 2.2, and 2.3).

API compatibility was retained, however the 2.0 release features a few nice new features and signifigant bugfixes

The new version can be installed from pypi using pip install -U imgstore.

Interested in Motif?

Motif is the first video and camera recording system designed for the experiments of modern scientists. It supports single and multiple synchronized camera scenarios, remote operation, high framerate and unlimited duration recording. It is always updated and has no single-user or other usage limitations.

If you are interested in aMotif system, please contact us for a quote or to see how Motif can solve your video recording needs.

A Multi-Camera High Throughput Worm Assay

Written on Wednesday October 17, 2018

We were approached by Andre Brown to create a custom imaging setup for high-throughput worm screening. Dr. Brown and his team are searching for novel neuroactive compounds using the nematode C. elegans. He explained:

"To screen a large number of drugs, it helps to image many worms in multiwell plates. Normal plate scanning microscopes don't help because we want to see behaviour changes that require looking at each worm for minutes so we have to image the worms in parallel. At the same time, we need enough resolution to extract a detailed behavioural fingerprint using Tierpsy Tracker. A six-camera array provides the pixel density and frame rate we need and opens the door to phenotypic screens for complex behaviours at an unprecedented scale."

The Solution

To meet the requirements we designed a solution using 6x 12 Megapixel cameras at 25 frames per second. To save on storage space, all video is compressed in real-time, at exceptional quality, before being saved to disk. The flexibility of Motif allows synchronized recording from all cameras, controlled from a single web-based user interface.

For an impression of the images possible with such a system, check out the interactive (mouse over / touch for zoom controls) viewer below.

unfortunately the generation of the visualisation above introducted some artifacts not present in the orignal video

Interested in Motif?

Motif is the first video and camera recording system designed for the experiments of modern scientists. It supports single and multiple synchronized camera scenarios, remote operation, high framerate and unlimited duration recording. It is always updated and has no single-user or other usage limitations.

If you are interested in a Motif system, please contact us for a quote or to see how Motif can solve your video recording needs.

OpenCV Conda Packages

Written on Friday May 25, 2018

At loopbio we maintain some linux packages for use with the conda package manager. These can replace the original packages present in the community-driven conda-forge channel, while retaining full compatibility with the rest of the packages in the conda-forge stack. They include some useful modifications that make them more suited to us, but that we find difficult to submit "upstream" for inclusion in the respective official packages.

Why might our packages be useful to you?

At the time of writing this note, we are actively maintaining three packages:

We have written a getting started with Conda guide here. If you are already familiar with conda then replacing your conda-forge packages with ours is a breeze. Using your command line:

# Before getting our conda packages, get a conda-forge based environment.
# For example, use conda-forge by default for all your environments.
conda config --add channels conda-forge

# install and pin ffmpeg GPL (including libx264)...
conda install 'loopbio::ffmpeg=*=*gpl*'

# ...or install and pin ffmpeg LGPL (without libx264)
conda install 'loopbio::ffmpeg=*=*lgpl*'

# install and pin libjpeg-turbo
# note, this is not needed for opencv to use libjpeg-turbo
conda install 'loopbio::libjpeg-turbo=1.5.90=noclob_prefixed_gcc48_*'

# install and pin opencv
conda install 'loopbio::opencv=3.4.3=*h6df427c*'

If you use these packages and find any problem, please let us know using each package issue tracker.

Example: controlling ffmpeg number of threads when used through OpenCV VideoCapture

We have added an environment variable OPENCV_FFMPEG_THREAD_COUNT that controls ffmpeg's thread_count, and a capture read-only property cv2.CAP_PROP_THREAD_COUNT that can be queried to get the number of threads used by a VideoCapture object. The reason why an environment variable is needed and the property is read only is that the number of threads is a property that needs to be set early in ffmpeg's lifecycle and should not really be modified once the video reader is open. Note that threading support actually depends on the codec used to encode the video (some codecs might, for example, ignore setting thread_count). At the moment we do not support changing the threading strategy type (usually one of slice or frame).

The following are a few functions that help controlling the number of threads used by ffmpeg when decoding a video via opencv VideoCapture objects.

  """OpenCV utils."""
  import contextlib
  import os
  import cv2
  import logging

  _log = logging.getLogger(__package__)

  def cv2_num_threads(num_threads):
      """Context manager to temporarily change the number of threads used by opencv."""
      old_num_threads = cv2.getNumThreads()

  # A string to request not to change the current value of an envvar
  USE_CURRENT_VALUE = object()

  def envvar(name, value=USE_CURRENT_VALUE):
      Context manager to temporarily change the value of an environment variable for the current process.

      Remember that some envvars only affects the process on startup (e.g. LD_LIBRARY_PATH).

      name : string
        The name of the environment value to modify.

      value : None, `cv2utils.USE_CURRENT_VALUE` or object; default "USE_CURRENT_VALUE"
        If `cv2utils.USE_CURRENT_VALUE`, the environment variable value is not modified whatsoever.
        If None, the environment variable value is temporarily removed, if it exists.
        Else, str(value) will be temporarily set as the value for the environment variable

      When a variable is not already set...
      >>> name = 'AN_ENVIRONMENT_VARIABLE'
      >>> with envvar(name, None):
      ...     print(os.environ.get(name))
      >>> with envvar(name, USE_CURRENT_VALUE):
      ...     print(os.environ.get(name))
      >>> with envvar(name, 42):
      ...     print(os.environ.get(name))
      >>> print(os.environ.get(name))

      When a variable is already set...
      >>> os.environ[name] = 'a_default_value'
      >>> with envvar(name, USE_CURRENT_VALUE):
      ...     print(os.environ.get(name))
      >>> with envvar(name, None):
      ...     print(os.environ.get(name))
      >>> print(os.environ.get(name))
      >>> with envvar(name, 42):
      ...     print(os.environ.get(name))
      >>> print(os.environ.get(name))
      if value is USE_CURRENT_VALUE:
      elif name not in os.environ:
          if value is not None:
              os.environ[name] = str(value)
              del os.environ[name]
          old_value = os.environ[name]
          if value is not None:
              os.environ[name] = str(value)
              del os.environ[name]
          os.environ[name] = old_value

  def ffmpeg_thread_count(thread_count=USE_CURRENT_VALUE):
      Context manager to temporarily change the number of threads requested by cv2.VideoCapture.

      This works manipulating global state, so this function is not thread safe. Take care
      if you instantiate capture objects with different thread_count concurrently.

      The actual behavior depends on the codec. Some codecs will honor thread_count,
      while others will not. You can always call `video_capture_thread_count(cap)`
      to check whether the concrete codec used does one thing or the other.

      Note that as of 2018/03, we only support changing the number of threads for decoding
      (i.e. VideoCapture, but not VideoWriter).

      thread_count : int or None or `cv2utils.USE_CURRENT_VALUE`, default USE

        * if None, then no change on the default behavior of opencv will happen
          on opencv 3.4.1 and linux, this means "the number of logical cores as reported
          by "sysconf(SC_NPROCESSORS_ONLN)" - which is a pretty aggresive setting in terms
          of resource consumption, specially in multiprocess applications,
          and might even be problematic if running with capped resources,
          like in a cgroups/container, under tasksel or numactl.

        * if an integer, set capture decoders to the specifiednumber of threads
          usually 0 means "auto", that is, let ffmpeg decide

        * if `cv2utils.USE_CURRENT_VALUE`, the current value of the environment
          variable OPENCV_FFMPEG_THREAD_COUNT is used (if undefined, then the default
          value given by opencv is used)
      return envvar(name='OPENCV_FFMPEG_THREAD_COUNT', value=thread_count)

  def cv2_supports_thread_count():
      """Returns True iff opencv has been built with support to expose ffmpeg thread_count."""
      return hasattr(cv2, 'CAP_PROP_THREAD_COUNT')

  def video_capture_thread_count(cap):
      Returns the number of threads used by a VideoCapture as reported by opencv.
      Returns None if the opencv build does not support this feature.
          # noinspection PyUnresolvedReferences
          return cap.get(cv2.CAP_PROP_THREAD_COUNT)
      except AttributeError:
          return None

  def open_video_capture(path,
      Returns a VideoCapture object for the specified path.

      path : string
        The path to a video source (file or device)

      num_threads : None, int or `cv2utils.USE_CURRENT_VALUE`, default None
        The number of threads used for decoding.
        If None, opencv defaults is used (number of logical cores in the system).
        If an int, the number of threads to use. Usually 0 means "auto", 1 "single-threaded"
        (but it might depend on the codec).

      fail_if_unsupported_num_threads : bool, default False
        If False, an warning is cast if num_threads is not None and setting the
        number of threads is unsupported either by opencv or the used codec.

        If True, a ValueError is raised in any of these two cases.

      backend : cv2 backend or None, default cv2.CAP_FFMPEG
        If provided, it will be used as preferred backend for opencv VidecCapture
      if num_threads is not None and not cv2_supports_thread_count():
          message = ('OpenCV does not support setting the number of threads to %r; '
                     'use loopbio build' % num_threads)
          if fail_if_unsupported_num_threads:
              raise ValueError(message)

      with ffmpeg_thread_count(num_threads):
          if backend is not None:
              cap = cv2.VideoCapture(path, backend)
              cap = cv2.VideoCapture(path)

      if cap is None or not cap.isOpened():
          raise IOError("OpenCV unable to open %s" % path)

      if num_threads is USE_CURRENT_VALUE:
              num_threads = float(os.environ['OPENCV_FFMPEG_THREAD_COUNT'])
          except (KeyError, TypeError):
              num_threads = None
      if num_threads is not None and num_threads != video_capture_thread_count(cap):
          message = 'OpenCV num_threads for decoder setting to %r ignored for %s' % (num_threads, path)
          if fail_if_unsupported_num_threads:
              raise ValueError(message)

      return cap

If you get these functions, you can open and read capture like this:

  # Do whatever you need
  if not cap.isOpened():
      raise Exception('Something is wrong and the capture is not open')
  retval, image = cap.read()

Hoping other people find these packages useful.

Getting Started With Conda

Written on Friday May 04, 2018

Here at loopbio gmbh we use and recommend the Python programming language. For image processing our primary choice is Python + OpenCV.

Customers often approach us and ask what stack we use and how we set up our environments. The short answer is: we use conda and have our own packages for OpenCV and FFmpeg.


In the following post, we will bravely explain how easy it is to set up a Conda environment for image processing using miniconda and our packages for OpenCV and a matched FFmpeg version on Linux (Ubuntu). If you are not familiar with the concept of Conda: Conda is a package manager and widely used in science, data analysis and machine learning, additionally, it is fairly easy and convenient to use.

If you are more interested in why we are using OpenCV, FFmpeg and Conda and what performance benefits you can expect from our packages please check out our other posts.

Install Miniconda

  1. Download the appropriate 3.X installer
  2. In your Terminal window, run: bash Miniconda3-latest-Linux-x86_64.sh
  3. Follow the prompts on the installer screens. If you are unsure about any setting, accept the defaults. You can change them later. To make the changes take effect, close and then re-open your Terminal window.
  4. Test your installation (a list of pacakages should be printed). conda list

More information is provided here

Setting up the environment

  # Before getting our conda packages, get a conda-forge based environment.
  # For example, use conda-forge by default for all your environments.
  conda config --add channels conda-forge

  # Create a new conda environment
  conda create -n loopbio

  # Source that environment
  source activate loopbio

  # install FFmpeg
  # install and pin ffmpeg GPL (including libx264)...
  conda install 'loopbio::ffmpeg=*=gpl*'

  # ...or install and pin ffmpeg LGPL (without libx264)
  conda install 'loopbio::ffmpeg=*=lgpl*'

  # install and pin opencv
  conda install 'loopbio::opencv=3.4.1'

Reading a video file

  # Make sure that the loopbio environment is activated
  source activate loopbio

  # Start Python

In Python

  import cv2
  cap = cv2.VideoCapture('Downloads/small.mp4')
  ret, frame = cap.read()
  print frame

Video I/O Part 2: Fast JPEG Decoding

Written on Tuesday April 17, 2018

In the previous installment of our series on Video I/O we threatened thorough benchmarks of video codecs. This series of blog posts is about ways to minimize delays in bringing video frames both to the browser and to video analysis programs, including training deep learning models from video data. In that post we showed plots like this one:

We used and will keep using what we called "exploded jpeg" as a baseline when talking about video compression, as encoding images as jpeg is, by far, the most commonly way to transport image data around in deep learning workloads. Because encoding and, more specially, decoding are important core operations for us in loopy, and also because we want to give ourselves a hard time trying to beat baselines, we strive to use the best possible software for encoding and decoding jpeg data.

So what is the fastest way to read and write jpeg images these days? And how can we get to use it in the most effective way? In this post we demonstrate that using libjpeg-turbo is the way to go, presenting the first independent benchmark (to our knowledge) of the newest jpeg turbo version and touching on a few related issues, from python bindings to libjpeg-turbo to accelerated python and libjpeg-turbo conda packages. So let's get started, shall we.

The Contenders

We are going to look across four dimensions here: libjpeg vs libjpeg-turbo, current stable version of libjpeg-turbo (1.5.3) vs the upcoming version (2.0), using libjpeg-turbo with different python wrappers, and using libjpeg-turbo with different parameters controlling the tradeoff between decoding speed and accuracy. On each round there will be a winner that gets to compete in the next one.

There is one main open source library used for jpeg encoding: libjpeg. And there is one main alternative to libjpeg for performance critical applications: libjpeg-turbo. Turbo is a fork of libjpeg where a lot of amazing optimization work has been done to accelerate it. Turbo works for many different computer architectures, and used to be a "drop-in" replacement for libjpeg. This stopped being true when libjpeg decided to adopt some non-standard techniques - perhaps hoping for them to become one day part of the jpeg standard. Turbo decided not to follow that path. In principle this means that there might be some non-standard jpeg images that turbo won't be able to decode[1]_. However, given the prevalence of turbo in mainstream software (for example, it is used in web browsers like firefox and chrome, and is a first class citizen in most linux distributions), it is unlikely these incompatibilities will be seen in the wild.

Having decided that libjpeg-turbo is to be used, we turn our attention to the python wrapper used on top of it. Our codebase has a strong pythonic aroma and therefore we are most interested on reading and writing jpegs from python code. Therefore we are using libjpeg, which is a C library, wrapped in python. We look here at two main wrappers: opencv, which we use as the go-to library for reading images, and a simple ctypes wrapper (modified from pyturbojpeg).

The simple ctypes wrapper exposes more libjpeg specific functionality from the wrapped C library such as faster but less accurate decoding modes. Usually these modes are deactivated by default, since they result in "less pleasant" images (for humans) in some circumstances. However certain algorithms might not notice these differences - for example tensorflow activates some of them by default under the (likely unchecked) premise that it won't matter for model performance.

The Benchmark

To measure how fast different versions of the libjpeg library can compress and decompres, we have used 23 different images from a public image compression benchmark dataset, some of our clients videos and even pictures of ourselves. We used these images at three different sizes, corresponding (without modifying the aspect ratio) to 480x270 (small), 960x540 (medium) and 1920x1080 (large) resolutions. We always used YUV420 as encoded color space and BGR as pixel format.

The following are three images from our benchmark dataset, at "medium" size, as originally presented to the codecs and after compression + decompression (with jpeg encoding quality set at 95 and using the fastest and less precise libjpeg-turbo decoding settings). Can you tell which one is the original and which one is the round-tripped version? (note, we have shuffled these a bit to make the challenge more interesting).

Loopy roundtripped image (faster decoding) Loopy original image C-elegans original image C-elegans roundtripped image (faster decoding) Cathedral original image Cathedral roundtripped image (faster decoding)

All data in this post is summarized results across all images, but it is important to note that when dealing with compression, results might vary substantially depending on the kind of images to be stored. In specific cases, such as when all images are similar, which might happen when storing video data as jpegs, it might be useful to select encoding/decoding parameters taylored to the data.

For each image and codec configuration we measured multiple times the round-trip encoding-decoding speed with randomized measurement order. We have checked that each roundtrip provides acceptable quality results using perceptual image comparison between the original image and the roundtripped one.

We have timed speed when using libjpeg and libjpeg turbo via python wrappers and subsequently the measurements always include some python specific costs - such as the time taken to allocate memory to hold the results. It is expected some speedups can be achieved by optimizing these wrappers memory usage strategies. We only measure speeds for image data already in RAM and that is expected to be "cache-warm", so these microbenchmarks represent a somehow idealized situation and should better be complemented with I/O and proper workload context. All measurements were made on a single core of an otherwise idle machine, sporting an intel i7-6850K and fairly slow RAM.

The Results

Encoding Speed

The following table shows average space savings for the benchmarked encode qualities. These are identical for all the libjpeg variants used and are compared against the space taken by the uncompressed image.

Encode Quality Average Space Savings
80 94.1 ± 2.8
95 87.2 ± 5.4
99 77.6 ± 8.4

Before showing our results summary, let us enumerate again the contenders:

  • opencv_without_turbo: opencv wrapping libjpeg 9b
  • opencv_with_turbo: opencv wrapping libjpeg-turbo stable (1.5.3)
  • turbo_stable: ctypes wrapper over libjpeg-turbo stable (1.5.3)
  • turbo_beta: ctypes wrapper over libjpeg-turbo 2.0 beta1 (1.5.90)
  • turbo_beta_fast_dct: like turbo_beta, activating "fast DCT" decoding for all passes
  • turbo_beta_fast_upsample: like turbo_beta, activating "fast upsampling" decoding
  • turbo_beta_fast_fast: like turbo_beta, activating both "fast DCT" and "fast upsampling"

The following plot shows how encoding speed varies across different compression qualities (you can show and hide contenders by clicking in the legend). We can see how libjpeg-turbo is a clear winner. opencv_without_turbo is doing the same job as its turbo counterpart opencv_with_turbo, just between 3 and 7 times slower. There is a second relatively large gap between using opencv or using directly turbo via ctypes, indicating that for high performance applications it would be worth to use more specific APIs. Finally, the upcoming version of libjpeg-turbo also brings a small performance bump worth pursuing.

Decoding Speed

The following plot shows decoding speed differences between our contenders, as a function of the image quality.

Again, turbo is just much faster than vanilla libjpeg, using the ctypes wrapper is much faster than using opencv, and using the newer version of turbo is worth it.

Three new candidates appear slightly on top of the speed ranking: turbo_beta_fast_dct, turbo_beta_fast_upsample and turbo_beta_fast_fast. These activate options that trade higher speed for less accurate (or less visually pleasant) approximations to decompression. They are deactivated by default in libjpeg-turbo, but other wrapper libraries, notably tensorflow, do activate them by default under the premise that machine learning should not be affected by the loss of accuracy. The same that our tests did not find any relevant difference on speed, they did not show any elevated loss on visual perception scores, so we do not have any strong opinion on activating them or not, we just think it is now mostly irrelevant.

Why is the ctypes wrapper faster than opencv? There probably are several reasons, but if one looks briefly to the opencv implementation, a clear suspect arises. With libjpeg-turbo you can specify wich pixel format the jpeg data is using (the jpeg standard is actually agnostic of which order do the color channels appear in the file) to avoid unneeded color space conversions. OpenCV instead goes a long way to non-optionally convert between RGB and BGR, probably to ensure that jpeg data is always RGB (which is a more common use) while uncompressed data is always BGR (a contract for opencv). Add to this that opencv barely expose some of the features of libjpeg and it does not have libjpeg-turbo specific bindings, our advice here would be to use a more specific wrapper to libjpeg-turbo.

Talking about wrappers, let's look at the last plot (for today). Here we show decoding speeds as a function of image size.

The larger the image, the faster we decompress. This is normal: there is some work that needs to be done before and after each decompression. The take home message here is, to our mind, to improve the wrappers to minimize constant performance overheads they might introduce. An obvious improvement is to use an already allocated (pinned if planning to use in GPUs) memory pool. This should prove beneficial, for example, when feeding minibatches to deep learning algorithms. A more creative improvement would be to stack several images together and compress them into the same jpeg buffer.

Note also that there are several features of turbo we have not explored here. An important example is support for partial decoding (decode a region (crop) of an image without doing all the work to decode the whole image), which was introduced in turbo recently (partially by google) and was then exposed in tensorflow. We have not actually found ourselves in need for these advanced features, but let the need come, we are happy to know we have our backs covered by a skilled community of people seeking our same goals: to get image compression and decompression times out of our relevant bottlenecks equations.

Speeding Up Your Code

TLDR; use our opencv and libjpeg-turbo conda packages

So how do one use libjpeg-turbo?. Well, as we mentioned, libjpeg-turbo is everywhere these days - so some software you run is probably already using it. If you are using Firefox or Chrome to read this post, it is very likely that jpeg images are being decompressed using turbo. If you use tensorflow to read your images, you are using turbo. On many linux distributions libjpeg-turbo is either the default package or can be installed to replace the vanilla libjpeg package. We are not very knowledgeable of what is the story with other platforms, but we suspect that libjpeg-turbo reach and importance extends to practically any platform where jpeg needs to be processed.

What if you use the conda package manager? In this case you might be a bit out of luck, because the two main package repositories (defaults and conda-forge) have moved to exclusively use libjpeg 9b in their stack. If you try to use a libjpeg-turbo package in a modern conda environment, chances are that you will bump into severe (segfaulty) problems. This is a bit of a disappointment given that conda is commonly used in the scientific, data analysis and machine learning fields these days.

But good news! if you are on linux your luck has changed - all you need to do is to use our opencv and libjpeg-turbo packages (which bring along our ffmpeg package). Because we use these packages in loopy we keep them in sync with the main conda channels and ready to be used by any conda user.

These packages avoid problems with parallel installations of libjpeg 9 and libjpeg-turbo, and offer other few goodies (like the ability to choose between GPL and non-GPL versions of ffmpeg or patches to control video decoding threading when done via opencv). The creation of these modified packages was not a small feature and will also be covered in a future blog post. In the meantime, you can just use any of these command lines to use the packages:

# before running this, you need conda-forge in your channels
conda config --add channels conda-forge

# this would get you our latest packages
conda install -c loopbio libjpeg-turbo opencv

# this would get you and pin our current packages (N.B. requires conda 4.4+)
conda install 'loopbio::opencv=3.4.1=*_2' 'loopbio::libjpeg-turbo=1.5.90=noclob_prefixed_gcc48_0'

Or add something like this to your environment specifications (note these are the exact software versions we used when benchmarking for this post):

name: jpegs-benchmark

  - loopbio
  - conda-forge
  - defaults


  # uncomment any of these to get the opencv / turbo combo you want

  # note that, at the moment of writing, these pins are not always respected
  # see: https://github.com/conda/conda/issues/6385

  # opencv compiled against turbo 2.0beta1
  - loopbio::opencv=3.4.1=*_2  # compiled against turbo 2.0beta1

  # opencv compiled against turbo 1.5.3
  # - loopbio::opencv=3.4.1=*_1

  # opencv compiled against libjpeg 9
  # - conda-forge::opencv=3.4.1

  # libjpeg-turbo 2.0beta1
  - loopbio::libjpeg-turbo=1.5.90=noclob_prefixed_gcc48_0

  # libjpeg-turbo 1.5.3
  # - loopbio::libjpeg-turbo=1.5.3

Finally, you can use these packages with our pyturbojpeg fork to achieve better performance than generic libjpeg wrappers like PIL or opencv. If you install both the turbo packages and our wrapper, you can easily compress and decompress jpeg data like this:

  from turbojpeg import Turbojpeg
  turbo = Turbojpeg()
  roundtrip = turbo.decode(turbo.encode(image))


If you need very fast jpeg encoding and decoding:

  1. Use turbo
  2. Use turbo with fast wrappers
  3. Use the latest version of turbo and decide for yourself if using faster modes for encoding and decoding is worth it.
  4. If you use conda then use our accelerated jpeg and opencv packages

We also think that any benchmark for, let's say, image minibatching for deep learning, should explicitly include a solution based on libjpeg-turbo as a contender.

Finally, use turbo also if you do not need any of these things as it is very easy to install (on linux) and will probably magically speed up many other things on your computer.

It is good to remember that open source software always needs a hand.

[1]From the developers libjpeg-turbo is currently under consideration for becoming an official ISO/ITU-T reference implementation. Furthermore the libjpeg 'SmartScale' extension has not been adopted and the likelihood of it being used even if it was - is low.

Video I/O Part 1: An Introduction to Video Compression

Written on Thursday April 05, 2018

Video compression is polarising topic and a key technology for us at Loopbio. In a series of blog posts, beginning with this one, we put our didatic engineers hat on and share our experiences on how to best leverage current compression technologies, with a focus on practical benchmarking of video storage solutions, ultimately seeking to find satisfactory trade offs between several parameters: video quality, disk space and speed.

The tentative TOC of this series is (links will be added as published);

  • Part 1 (this one): an introduction to video compression for novices
  • Part 2: fast jpeg decoding
  • Part R: benchmarking lossless compression codecs
  • Part E: benchmarking lossy compression codecs
  • Part A: notes on high resolution videos
  • Part D: reading videos for deep learning training (minibatching take)
  • Part I: our conda packages (or how to decompress images faster)
  • Part L: benchmarking tricks and tips
  • Part Y: checking correctness of video seeking

Our company mission is to deliver best in class, easy to use, video analysis solutions. For that we have created loopy, a platform for working with arbitrarily large collections of videos. It lets the users, from the comfort of their browsers, explore, annotate and organize their videos. loopy also offers state of the art analysis tools, from 3D vision to deep-learning powered tracking.

To achieve our mission, loopy needs to be user friendly when used interactively. It must also cleverly use possibly scarce hardware resources, specially when running computationally intensive tasks. Our infrastructure needs are also complicated by the fact that loopy is offered in both an online subscription-based version, and an on-premise managed server version.

Therefore, delays bringing video frames both to the browser and to the analysis programs must be minimized. In particular, as we will see, we should be able to accurately read from arbitrary video positions in as little time and with as little computation as possible, and with absolute accuracy.

A concise primer on video compression

Let us first start with a reading recommendation. This introduction to digital video and references therein, which you can read from the comfort of a gym machine - if you are into these kind of things - is not to be missed. It contains an amazingly visual description of the basics of video coding and related topics, from the human visual system to codec-wars, touching on many of the topics we will speak about in the next few paragraphs.

Basic terminology

A video is just a sequence of images (frames), happening at a regular (generally but not required) frequency in time (framerate), usually measured in frames per second (FPS). When stored digitally, these images are usually composed of a number of pixels, "dots with a color", organized in a grid. The video has therefore a resolution that indicates the shape of this grid, and which is usually indicated with width (number of pixels per row) and height (number of pixels per column).

How colors are actually represented and their spectrum is given by the so called color space. Each individual color is represented by a number/(s) indicating the intensity of certain color components. Videos have therefore also a color depth, that is measured in bits and indicates how much information is needed to represent the color of a single pixel. Usual color depths are 8 (for grayscale images), 24 (for "true color" images) and 48 (for "deep color" images).

A video is a many-D matrix


When stored digitally, an uncompressed video needs `width * height * colordepth * framerate * duration` bits. How much is that? As an example, one early video that a client uploaded to loopy had a resolution of 1920x1080, 24 bits color depth and a fast framerate of 120fps. If uncompressed, this video would need 5 971 968 000 bits per second (this is know as bitrate). In other words, a minute of such video would use up around 42GB, or in other words, had we stored these videos raw, we could only have been able to keep around 100 minutes. Our client has around 145 hours of beautiful fish schools footage recorded under the Red See, so there is no way that would work.

Obviously no one uses raw video when storing or transmitting digital video. Our client had around 145 hours of footage, but it was taking only slightly less than 7TB (instead of 2835TB!). This is so because the videos were compressed.

Video compression takes advantages of spatial redundancies (regions of a single frame with a lot or repetition), temporal redundancies (consecutive frames tend to look very much alike) and the human visual perception particularities (for example, we can distinguish better between bright and dark than between shades of colors) to make storing and transmitting videos leaner, while still keeping good video quality.

Example frame from a video Example frame from a video

The two previous images (by courtesy of Simon Gingins) are two consecutive frames from a video of the mentioned collection. On each image the background is mostly blue (spatial redundancy) and the difference between them is minimal (if you can tell, fishes have moved slightly).

The programs that compress and uncompress videos are called video codecs (a portmanteau of coder-decoder). While the techniques used on each of them are highly related, there are many different codecs with different characteristics. Different codecs may generate larger or smaller videos, can compress and decompress at different speeds and/or using different amount of computer resources, can allow better or worse random access patterns, and can produce higher or lower quality videos. In this series of blog posts, we will be interested in finding out which video codecs are competitive, across many parameters, to fill different roles inside our platform.

Compression quality

A very important distinction between codecs is the quality of the output video, that is, how faithfully the compressed video represents the original material. Codecs can trade size for quality. When aiming at reduced size, codecs can use tricks that result in video artifacts. We probably are all used to these kinds of obvious errors when, for example, viewing streaming movies or looking at photos. While watching netflix, did you ever realized how pixelated that super-red cape from SuperWoman looked like?

Obvious macroblocking errors Obvious macroblocking errors

An important dividing line here is between lossless codecs, codecs that ensure no loss of quality, and lossy codecs, codecs that do not make such guarantee. Usually lossy codecs can be made to produce good quality video (almost perceptually lossless) at a much higher space savings than lossless codecs by exploiting the redundancies explained above.

Note: for this post and in general for this series, lossless codecs are assumed to be operating under their native colorspace, we are thus ignoring any numerical effects from colorspace conversion.

Lossless codecs receive much less mainstream attention than lossy ones, despite them serving many important roles (in loopy and elsewhere). For example, they are usually tasked with compressing video for archival or editing purposes. We will in this series make lossless codecs an integral part of our benchmarks, assessing how well they fare when compressing interesting parts of larger video collections.

In particular, we will investigate their suitability when storing short annotated regions of videos (clips) and frames (crops) for the purposes of batching/training a neural network for object detection (discussed in a future post). In this workload, loopy needs fast-access caches to concrete frames (in particular, frames that are annotated) in order to feed them to our machine learning algorithms.

Types of frames and video seeking

A final video-codec concept we wish to introduce is that of frame type. As we have said, video codecs can exploit intra-frame redundancy, that is, pixel redundancy within the same image, and inter-frame redundancy, that is, smooth changes between consecutive frames. Naively simplified, inter-frame compression uses the difference between frames to encode the video.

Videos compressed this way can contain up to three types of frames. Intra frames (I) are self contained and can be decoded without refering to any other frame. On the other hand, Predictive (P) frames require previous frames to be decoded first, while Bi-directional predictive (B) frames require both previous and posterior frames.

Video frame types

The main use for video is sequential playback. Modern video codecs can employ quite complex combinations of I, P and B frames, and sort the data for frames in an arbitrary order. Usually I frames are much less frequent, because they consume more space. This is one of the reasons why seeking, that is, reading an arbitrary frame from a video, is a slow operation. When requesting an arbitrary frame, it is likely we will need to decode many other frames to get the result.

To the best of our knowledge, there are no codecs that are specifically designed for fast seeking (that is, fast random frame retrieval). Seeking is a important operation for many tasks that loopy needs to do routinely, so when too slow, it might affect negatively the user experience. loopy often needs to display arbitrary segments of a video, often accessed in a pseudo-random order (for deep learning or by the user), with the video sometimes on slow underlying storage.

That is why we will also pay attention to seeking performance of video codecs, which is not a topic usually covered by video benchmarks.

Video collections at loopbio

Our company currently focus on servicing the life-sciences and, in particular, the behavioral research community. Because of that, we get to analyze some particular kinds of videos. The most common video collection we help analyzing contains hours of high-resolution, fast-framerate animal behavior recordings. Usually these come from either still cameras - both in natural environments and behavioral arenas, over and underwater - or aerial (drone) cameras. A recurrent characteristic of these videos is a fairly constant background that, in many cases, is also very simple.

We have expended a few fun minutes browsing youtube for publicly available examples of these kind of videos, and have selected just a few:

C-elegans fish-mirror Drone over Niamey The Manakin dance A beautiful undersea video

We will be using these videos to illustrate a few concepts about video compression and to benchmark video compression alternatives in later posts. To finish on an inviting note, we will now show a couple of results from our benchmarks on lossless codecs. We will provide codec descriptions and describe in detail our benchmarking methodology in later installments of this series.

Codec Selection Matters: Video Encoding Benchmark

How fast and how much do lossless codecs compress our benchmarking video suite? We asked a few lossless codecs to re-compress the original videos at different resolutions and we measured both quantities - disk savings and speed. We summarize some of our findings in the following plot (click on the plot legend to show/hide series):

A few codecs contend. ffv1, as we used it, is an intra-frame only codec geared towards a good trade off between compression speed and space savings. huffyuv/ffvhuff is a simple intra-frame only codec geared towards fast decompression. We used ffmpeg to encode for both ffv1 and ffvhuff. H.264 (encoded using ffmpeg bindings to libx264) is a fairly sophisticated codec, with many nuts and bolts that can be tweaked as needed; here we use its lossless mode. Each codec is parametrized in different ways (see and click the legend).

We also include a lossy baseline - and it is quite a strong baseline - which will serve as reference for speed during our benchmarks. We call it "exploded-jpg" and it essentially works by storing the video by saving each frame individually to a good quality jpeg-compressed file - so it is not a video-codec per se, although there are codecs (mjpeg) that work in an analogous way.

Note the wide range of performances. In the vertical axis we plot the space saved by compressing the video as a percentage of what would have taken to store it uncompressed. Higher is better and so H.264 is the winner. In the horizontal axis we plot the speed-up relative to the baseline. Again, the higher the better. So H.264, properly parametrized, is a win win here.

Why do we have such broad error bars? A statistician would probably start thinking about small data effects. While some of it might be true, here it does not really tell the whole story, or even the most relevant part of the story. Instead, this plot is a summary of writing workloads for different video types at different resolutions. Codecs can behave quite differently depending on those conditions. To us, these large error bars send this message: while having good defaults informed by relevant data is important, sometimes it can repay to engineer for the very particulars of an application. Obviously this is easier if one counts with trustworthy benchmarking tools

An important note about the baseline speed. It is generally much slower (2x to 4x) than the video codecs. Truth to be told, we did not parallelize it. That is, while the video codecs were free to use as many resources from our system as they wanted to - and all of them made the computer sweat a lot - we only allowed the baseline to use 1 sad core. Note that, usually, parallelism needs to be taken into account explicitly in a benchmark. When in the gym, I could only read video compression tutorials when taking it easy over the static bike. If I would work more intensely, then I would have lacked the resources to also do anything else at the same time.

What we did use is for the baseline is a highly optimized JPEG codec. This is too important to be overlooked, as we will discuss soon in this series.

While writing speed is very important for video acquisition systems like our own Motif, for loopy writing speed only has a small relative importance. For this reason, from now on we will focus our efforts on thorough benchmarking "write-once, read-many times" workloads.

How slow is seeking with lossless codecs?

On our second and concluding for now result, we plotted relative speed of retrieving a single random frame from a video compressed with the codecs used above against the baseline.

Video codecs are, in the best case, an order of magnitude slower when serving arbitrary video frames. This result holds no matter the underlying storage system (let it be a SSD, a spin disk or a network file system). That would immediately disqualify them for many machine learning workloads, were a program learns to perform some tasks by being exposed to examples of such task in random order. Random data reading must be as much as possible out of the way in a workload that contains many other heavy computation components. Therefore, the difference between a high and a low speed solution is the difference between speedy result delivery and resource infra-utilization leading to delays that can be counted in days or weeks.

Videos encoded as JPEG images, our baseline, is in fact the most common way in which video training data is stored. We are exploring if there is a way to use more efficient access patterns so that video codecs can somehow become an option in this arena, bringing some further benefits, like higher image quality, better introspectability and web playability, with them.

See you soon

If you are interested in our comprehensive video reading benchmarks follow us on twitter in @loopbiogmbh.