Skip to content

scanner-research/vgrid

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

VGrid: video metadata visualization in Javascript

Build Status

VGrid is a Javascript library for visualizing spatiotemporal metadata (e.g. bounding boxes, pose keypoints) on video using spatiotemporal intervals. For example, our group uses VGrid to inspect and debug computer vision models.

VGrid is a standalone JS library, so you can embed it any web page. See the vgrid_jupyter package for Jupyter integration. VGrid also has a Python library for building visualizations and exporteing them to JSON.

Installation

VGrid has two API components: the main (required) Javascript API for rendering the visualizing component, and an (optional) Python API for creating VGrid inputs. The Python API is useful if you have a Jupyter notebook or a Python server with VGrid on the frontend. We distribute prebuilt JS and Python packages, but you can also build from source.

Javascript API

VGrid must be installed in the context of a Javascript application using the npm package structure. You must have npm installed. VGrid has react, react-dom, mobx, and mobx-react as peer dependencies, so you must have those npm packages already installed in your Javascript application, e.g.

npm install --save react react-dom mobx mobx-react

Then you can install VGrid:

npm install --save @wcrichto/vgrid

Python API

VGrid requires Python 3.5 or greater.

pip3 install vgridpy

Example usage

Assume you have a video test.mp4 in your current directory. This is an example that shows a block with a single interval from time 0s to 10s for the video. First, you need to launch a local file server that can serve the video file to a webpage. For example, you can launch a server in the same directory as test.mp4 on port 8000 by running:

python3 -m http.server

Javascript only

If you're only running Javascript, then you will need to construct the video metadata and interval data using the Javascript API.

import ReactDOM from 'react-dom';
import {VGrid, Table, Database, IntervalSet, Interval, Bounds, BoundingBox, SpatialType_Bbox} from '@wcrichto/vgrid';
import '@wcrichto/vgrid/dist/vgrid.css';

// Setup intervals
let interval_blocks = [{
  video_id: 0,
  interval_sets: [{
    name: 'test',
    interval_set: new IntervalSet([new Interval(
      new Bounds(0, 10),
      {
        spatial_type: new SpatialType_Bbox(),
        metadata: {}
      }
    )])
  }]
}];

// Associate video IDs with metadata
let database = new Database(
  [new Table(
    'videos',
    [{id: 0, path: 'http://localhost:8000/test.mp4', num_frames: 1000,
      width: 640, height: 480, fps: 29.97}])]);

// Global component settings
let settings = {
  show_timeline: true
};

// Run code when user provides labeling input
let label_callback = (label_state) => {
  console.log(label_state.blocks_selected);
};

// Render React component into a <div id="#vgrid"></div>
ReactDOM.render(
  <VGrid interval_blocks={interval_blocks} database={database}
         settings={settings} label_callback={label_callback} />,
  document.getElementById('vgrid'));

Javascript and Python

You can use the Python API to build the video metadata and interval data, then send the JSON string to the frontend. Python:

from rekall import Interval, IntervalSet, Bounds3D
from vgrid import VGridSpec, IntervalBlock, NamedIntervalSet, VideoMetadata

video_id = 1
video = VideoMetadata(path='test.mp4', id=video_id)
intervals = IntervalSet([Interval(Bounds3D(0, 10))])
interval_blocks = [
  IntervalBlock(
    video_id=1,
    interval_sets=[
      NamedIntervalSet(name='test', interval_set=intervals)
    ])
]

vgrid_spec = VGridSpec(
  video_meta=[video],
  interval_blocks=interval_blocks,
  show_timeline=False)

# Send vgrid_spec.to_json() and settings to the frontend somehow

And Javascript:

import ReactDOM from 'react-dom';
import {VGrid, Database, interval_blocks_from_json} from '@wcrichto/vgrid';

// Fetch the JSON somehow
fetch_json_somehow(function(vgrid_spec) {
  // Convert JSON into corresponding Javascript objects
  let {interval_blocks, database, settings} = vgrid_spec;
  let database = Database.from_json(database);
  this.interval_blocks = interval_blocks_from_json(interval_blocks);

  // Run code when user provides labeling input
  let label_callback = (label_state) => {
    console.log(label_state.blocks_selected);
  };

  // Render React component into a <div id="#vgrid"></div>
  ReactDOM.render(
    <VGrid interval_blocks={interval_blocks} database={database}
           settings={settings} label_callback={label_callback} />,
    document.getElementById('vgrid'));
});

Examples

To see more examples in Javascript and Python, check out the examples folder!

API

VGrid provides a means of visualizing spatiotemporal data on video. We represent spatiotemporal data using intervals and interval sets as defined in the Rekall library.

Interval blocks

VGrid is a grid of interval blocks, where each block contains multiple sets of related intervals within a single video. For example, let's say I have four videos, and each video has three sequences, and each sequence has three interval sets (face bounding boxes, pose keypoints, and captions). That can be represented as a grid of twelve interval blocks with each block containing three interval sets.

VGrid provides a React component as shown in the examples above. The parameters are documented in the VGridProps type.

The first required parameter is a list of IntervalBlock specifying the spatiotemporal metadata. Each block contains a video ID and a list of NamedIntervalSet.

The video ID matches up with the second parameter, a Database containing metadata about videos and other entities referenced in the intervals. The database module documents the required fields for different tables, e.g. for videos and for categories.

The Python API provides an equivalent interface. It also provides a number of higher-level visualization formats that generate interval blocks from IntervalSetMapping objects. For example, the VideoBlockFormat shows one interval block per video:

from rekall import Interval, IntervalSet, Bounds3D, IntervalSetMapping
from vgrid import VGridSpec, VideoMetadata, VideoBlockFormat

video_id = 1
video = VideoMetadata(path='test.mp4', id=video_id)
intervals = IntervalSet([Interval(Bounds3D(0, 10))])
interval_map = IntervalSetMapping({video_id: intervals})

vgrid_spec = VGridSpec(
  video_meta=[video],
  vis_format=VideoBlockFormat([('test', interval_map)]),
  show_timeline=False)

Spatial types

A spatiotemporal interval is a 3-dimensional rectangular prism in time and x/y of the video. See the Rekall documentation for an extended discussion of how to think about this data representation.

To know how to draw an interval on a video, VGrid uses intervals annotated with a SpatialType. The default way to draw a spatiotemporal interval is with a box (SpatialType_Bbox) which corresponds to a rendering function for that draw type (BoundingBoxView).

Interval annotations like spatial type are contained within an interval's payload. Specifically, intervals must have a VData payload type containing a spatial type and a dictionary of Metadata. For example, in Javascript, we can explicitly create such an interval like this:

import {Interval, Bounds, SpatialType_Bbox} from '@wcrichto/vgrid';

let itvl = new Interval(
  new Bounds(0, 1), // from time 0s to time 1s
  {
    spatial_type: new SpatialType_Bbox(),
    metadata: {}
  });

Similarly in the Python API:

from rekall import Interval, Bounds3D
from vgrid import SpatialType_Bbox

intvl = Interval(
    Bounds3D(0, 1),
    {
        'spatial_type': SpatialType_Bbox(),
        'metadata': {}
    }
)

VGrid also provides a "builder" interface in the Python API to simplify the creation of intervals. See the module comments vblocks_builder.py for documentation and examples. Currently, VGrid supports the following draw types:

Bounding box

Bounding box: draws a rectangle over a video, default for all intervals.

Keypoints

Keypoints: draws a graph of nodes and edges over a video, useful e.g. for face or pose keypoints.

Captions

Captions: shows a box of captions beneath the video, does not draw on the video.

Metadata

Draw type describes the "core" of how an interval should be displayed, but it's also useful to have ancillary metadata about an interval. For example, an interval could be part of some category (e.g. comes from a video of type X, is an object of type Y). VGrid requires intervals to be annotated with a dictionary of Metadata. For example:

import {Interval, Bounds, Metadata_Generic} from '@wcrichto/vgrid';

let itvl = new Interval(
  new Bounds(0, 1), // from time 0s to time 1s
  {
    metadata: {hello: new Metadata_Generic('world')}
  });

When this interval is drawn, the the "hello: world" metadata will appear adjacent to the video. Currently, VGrid supports the following metadata types:

Generic

Generic: most basic kind of metadata. Will be shown just as a string.

Categorical

Categorical: metadata that is one value of a category, e.g. object type. Category name and type must be associated with a corresponding category table in the database.

Labeling

VGrid provides some capabilities for users to modify or create intervals. All modifications are tracked in the LabelState, which is passed to the top-level label_callback whenever the label state changes. Currently, VGrid supports the following label operations:

Categorizing blocks

Individual interval blocks can be categorized, currently as "Positive" or "Negative" (see LabelState.blocks_selected and BlockSelectType). The user labels blocks by clicking "s" or "x" while hovering over a block.

Creating temporal intervals

Modifying or creating individual intervals is reflected in the BlockLabelState (see also LabelState.block_labels). New temporal intervals (not spatial) can be created by clicking "i" while playing through a video.