V- Modal AI: Visual Video / Image Search - Flutter Video SDK
Dart android-sdkflutterflutter-appflutter-packageflutter-sdkflutter-videoflutter-video-editingflutter-video-playerflutter-video-sdkflutter-videosflutter-videosdkmobile-videomobile-video-editor-sdkmobile-video-sdkvideo-editingvideo-sdkvideo-searchvideo-search-enginevideo-streaming
import 'package:flutter/material.dart';
import 'theme.dart';
/// One bundled clip, with the attribution it must carry.
@immutable
class FootageClip {
const FootageClip({
required this.asset,
required this.label,
required this.credit,
});
final String asset;
/// Shown in the UI, e.g. a camera name or a clip title.
final String label;
/// Source and licence. Every clip in this app is CC0 stock.
final String credit;
String get fileName => asset.split('/').last;
}
/// A complete demo: the story, the footage, the queries and the look.
@immutable
class Scene {
const Scene({
required this.id,
required this.tabLabel,
required this.tabIcon,
required this.title,
required this.strapline,
required this.sourceLabel,
required this.suggestedCollection,
required this.footage,
this.exampleQueries = const <String>[],
required this.searchHint,
this.relevanceWindow = 0.03,
this.relevanceCeiling = 0.87,
required this.explainer,
required this.steps,
required this.palette,
});
final String id;
final String tabLabel;
final IconData tabIcon;
final String title;
final String strapline;
/// Label above the footage card: the camera, or the reel.
final String sourceLabel;
/// Suggested collection name, offered in the setup sheet. The one actually
/// used is whatever the user types, since a collection belongs to their key.
final String suggestedCollection;
final List<FootageClip> footage;
/// Tappable queries, where the demo benefits from suggesting a few. Each one
/// must have an obvious hit in the bundled clips, or the demo undersells the
/// SDK. Leave empty for a scene that reads better without them.
final List<String> exampleQueries;
final String searchHint;
/// How far past the best frame's distance to keep others. Tight suits one
/// camera, where the number of moments should mean something. Loose suits an
/// archive, where the question is which clips hold the shot at all and
/// cutting to the single best clip answers the wrong question.
final double relevanceWindow;
/// Absolute distance past which nothing is shown, whatever the best match
/// was. Stops an absurd query returning a confident-looking page.
final double relevanceCeiling;
/// What this demo is for, in two sentences. Shown in the info panel so a
/// screenshot of it explains the tab on its own.
final String explainer;
/// Three short steps shown before the first search, so the empty screen
/// explains the demo instead of looking unfinished.
final List<String> steps;
final Palette palette;
}
/// Traffic camera: an hour of junction footage, find the moment that matters.
const Scene trafficScene = Scene(
id: 'traffic',
tabLabel: 'Traffic',
tabIcon: Icons.videocam_outlined,
title: 'Traffic camera',
strapline: 'Describe the moment. Go straight to the frame.',
sourceLabel: 'Camera feed',
suggestedCollection: 'traffic_camera',
footage: <FootageClip>[
FootageClip(
asset: 'assets/footage/traffic_01.mp4',
label: 'CAM 04 — CITY JUNCTION',
credit: 'Pexels video 15643210, free to use — see README',
),
],
// Verified against the bundled clip: each of these returns frames that
// visibly contain what was asked for. Short object nouns do not work well —
// describe the moment, not the object.
exampleQueries: <String>[
'pedestrians on a zebra crossing',
'a bus waiting at the junction',
'a taxi',
],
searchHint: 'Describe what you are looking for',
explainer: 'One camera. Hours of footage. Find the moment.',
steps: <String>['Upload a clip', 'Describe the moment', 'Open the frame'],
palette: consolePalette,
);
/// Archive: many clips, and no idea which one holds the shot you want.
const Scene archiveScene = Scene(
id: 'archive',
tabLabel: 'Archive',
tabIcon: Icons.video_library_outlined,
title: 'Find the correct video',
strapline: 'Which clip has the shot you need?',
sourceLabel: 'Archive',
suggestedCollection: 'media_archive',
// Several clips indexed together. Results say which clip each frame is from,
// which is the job here: find the clip, not the moment.
// The junction clip is the same file the traffic tab uses, indexed here as
// one of several rather than on its own.
footage: <FootageClip>[
FootageClip(
asset: 'assets/footage/traffic_01.mp4',
label: 'City junction',
credit: 'Pexels video 15643210, free to use — see README',
),
FootageClip(
asset: 'assets/footage/ocean_01.mp4',
label: 'Ocean',
credit: 'Pexels, free to use — see README',
),
FootageClip(
asset: 'assets/footage/peppers_01.mp4',
label: 'Cutting peppers',
credit: 'Pexels, free to use — see README',
),
FootageClip(
asset: 'assets/footage/bird_01.mp4',
label: 'Bird',
credit: 'Pexels, free to use — see README',
),
FootageClip(
asset: 'assets/footage/cat_01.mp4',
label: 'Cat',
credit: 'Pexels, free to use — see README',
),
],
searchHint: 'Describe a shot in the archive',
// Measured across these five clips. Clips that genuinely hold the subject
// score 0.70 to 0.87; clips that do not start at 0.885. So the ceiling sits
// at 0.88: it keeps the ocean clip for "clouds" at 0.870, and drops the
// junction clip for "animal" at 0.906, which a looser 0.92 let through.
relevanceWindow: 0.15,
relevanceCeiling: 0.88,
explainer: 'Many clips. One search. Find which one.',
steps: <String>[
'Upload your clips',
'Describe it, or use a picture',
'See which clip has it',
],
palette: galleryPalette,
);
const List<Scene> scenes = <Scene>[trafficScene, archiveScene];
/// Pictures shipped for trying the image search, published to the device's
/// files alongside the footage so the picker can see them.
const List<String> queryImages = <String>[
'assets/queries/bird.jpg',
'assets/queries/sunset.jpg',
'assets/queries/clouds.jpg',
];