To explain how Scanner is used, let’s walk through a simple example that reads every third frame from a video, resizes the frames, and then creates a new video from the sequence of resized frames.


This Quickstart walks you through a very basic Scanner application that downsamples a video in space and time. Once you are done with this guide, check out the examples directory for more useful applications, such as using Tensorflow for detecting objects in all frames of a video and Caffe for face detection.

To run the code discussed here, install Scanner (Installation). Then from the top-level Scanner directory, run:

cd examples/apps/quickstart
wget https://storage.googleapis.com/scanner-data/public/sample-clip.mp4
python3 main.py

After main.py exits, you should now have a resized version of sample-clip.mp4 named sample-clip-resized.mp4 in the current directory. Let’s see how that happened by looking inside main.py.

Starting up Scanner

The first step in any Scanner program is to create a Database object. The Database object manages videos or other data that you have may have stored from data processing you’ve done in the past. The Database object also provides the API to construct and execute new video processing jobs.

from scannerpy import Database, Job

db = Database()

Ingesting a video into the Database

Scanner is designed to provide fast access to frames in videos, even under random access patterns. In order to provide this functionality, Scanner first needs to analyze the video to build an index on the video. For example, given an mp4 video named example.mp4, we can ingest this video as follow:

db.ingest_videos([('table_name', 'example.mp4')])

Scanner analyzes the file to build the index and creates a Table for that video in the Scanner database called table_name. You can see the contents of the database by running:

>>> print(db.summarize())
                      ** TABLES **
ID | Name       | # rows | Columns      | Committed
0  | table_name | 360    | index, frame | true

By default, ingest copies the video data into the Scanner database (located at ~/.scanner/db by default). However, Scanner can also read videos without copying them using the inplace flag.

db.ingest_videos([('table_name', 'example.mp4')], inplace=True)

This still builds the index for accessing the video but avoids copying the files into the database.

Defining a Computation Graph

Now we can tell Scanner how to process the video by constructing a computation graph. A computation graph is a graph of input nodes (Sources), function nodes (Ops), and output nodes (Sinks). Sources can read data from the Scanner database (such as the table we ingested above) or from other sources of data, such as the filesystem or a SQL database. Ops represent functions that transform their inputs into new outputs. Sinks, like Sources, write data to the database or to other forms of persistent storage.

Let’s define a computation graph to read frames from the database, select every third frame, resize them to 640 x 480 resolution, and then save them back to a new database table. First, we’ll create a Source that reads from a column in a table:

frame = db.sources.FrameColumn()

But wait a second, we didn’t tell the Source the table and column it should read from. What’s going on? Since it’s fairly typical to use the same computation graph to process a collection of videos at once, Scanner adopts a “binding model” that lets the user define a computation graph up front and then later “bind” different videos to the inputs. We’ll see this in action in the Defining Jobs section.

The frame object returned by the Source represents the stream of frames that are stored in the table, and we’ll use it as the input to the next operation:

sampled_frame = db.streams.Stride(input=frame, stride=3) # Select every third frame

This is where we select only every third frame from the stream of frames we read from the Source. This comes from a special class of ops (from db.streams) that can change the size of a stream, as opposed to transforming inputs to outputs 1-to-1.

We then process the sampled frames by instantiating a Resize Op that will resize the frames in the frame stream to 640 x 480:

resized = db.ops.Resize(frame=sampled_frame, width=640, height=480)

This Op returns a new stream of frames which we call resized. The Resize Op is one of the collection of built-in Ops in the Scanner Standard Library. (You can learn how to write your own Ops by following the Tutorials.)

Finally, we write these resized frames to a column called ‘frame’ in a new table by passing them into a column Sink:

output_frame = db.sinks.Column(columns={'frame': resized})

Putting it all together, we have:

frame = db.sources.FrameColumn()
sampled_frame = db.streams.Stride(input=frame, stride=3)
resized = db.ops.Resize(frame=sampled_frame, width=640, height=480)
output_frame = db.sinks.Column(columns={'frame': resized})

At this point, we have defined a computation graph that describes the computation to run, but we haven’t yet told Scanner to execute the graph.

Defining Jobs

As alluded to in Defining a Computation Graph, we need to tell Scanner which table we should read and which table we should write to before executing the computation graph. We can perform this “binding” of arguments to graph nodes using a Job:

job = Job(op_args={
    frame: db.table('table_name').column('frame'),
    output_frame: 'resized_table'

Here, we say that the FrameColumn indicated by frame should read from the column frame in the table "table_name", and that the output table indicated by output_frame should be called "resized_table".

Running a Job

Now we can run the computation graph over the video we ingested. This is done by simply calling run on the database object, specifying the jobs and outputs that we are interested in:

output_tables = db.run(output=output_frame, jobs=[job])

This call will block until Scanner has finished processing the job. You should see a progress bar while Scanner is executing the computation graph. Once the job are done, run returns the newly computed tables, here shown as output_tables.

Reading the results of a Job

We can directly read the results of job we just ran in the Python code by querying the frame column on the table resized_table:

for resized_frame in db.table('resized_table').column('frame').load():

Video frames are returned as numpy arrays. Here we are printing out the shape of the frame, which should have a width of 640 and height of 480.

Exporting to mp4

We can also directly save the frame column as an mp4 file by calling save_mp4 on the frame column:


After this call returns, an mp4 video should be saved to the current working directory called resized-video.mp4 that consists of the resized frames that we generated.

That’s a complete Scanner pipeline! If you’d like to learn about process multiple jobs, keep reading! Otherwise, to learn more about the features of Scanner, either follow the Interactive Jupyter Walkthrough or go through the extended Tutorials.

Processing multiple videos

Now let’s say that we have a directory of videos we want to process, instead of just a single one as above. To see the multiple video code in action, run the following commands from the quickstart app directoroy:

wget https://storage.googleapis.com/scanner-data/public/sample-clip-1.mp4
wget https://storage.googleapis.com/scanner-data/public/sample-clip-2.mp4
wget https://storage.googleapis.com/scanner-data/public/sample-clip-3.mp4
python3 main-multi-video.py

After main-multi-video.py exits, you should now have a resized version of each of the downloaded videos named sample-clip-%d-resized.mp4 in the current directory, where %d is replaced with the number of the video.

There are two places in the code that need to change to process multiple videos. Let’s look at those pieces of code inside main-multi-video.py now.

Ingesting multiple videos

The first change is that we need to ingest all of our videos. This means changing our call to ingest_videos to take a list of three tuples, instead of just one:

videos_to_process = [
    ('sample-clip-1', 'sample-clip-1.mp4'),
    ('sample-clip-2', 'sample-clip-2.mp4'),
    ('sample-clip-3', 'sample-clip-3.mp4')

# Ingest the videos into the database

Now we have three tables that are ready to be processed!

Defining and executing multiple Jobs

The second change is to define multiple jobs, one for each video that we want to process.

jobs = []
for table_name, _ in videos_to_process:
    job = Job(op_args={
        frame: db.table(table_name).column('frame'),
        output_frame: 'resized-{:s}'.format(table_name)

Now we can process these multiple jobs at the same time using run:

output_tables = db.run(output=output_frame, jobs=jobs)

Like before, this call will block until Scanner has finished processing all the jobs. You should see a progress bar while Scanner is executing the computation graph as before. Once the jobs are done, run returns the newly computed tables, here shown as output_tables.