LeanSceneScene detection demo

LeanScene finds where each shot (one unbroken take from a single camera) ends and the next begins: hard cuts, and gradual transitions such as dissolves and fades. Drop a video and see every one on a timeline, with the frame before and after.

Loading model…

Higher F1 than TransNet V2 on all three of its benchmarks, with 115× fewer operations per frame, in a 1.6 MB model.

Accuracy

F1 on the three standard shot-boundary benchmarks, scored the same way for both models

BBC Planet Earth
97.4 +1.3
ClipShots
79.1 +1.4
RAI
93.7 +1.1
7580859095100
LeanSceneTransNet V2F1, mean of 5 training runs; axis starts at 75

Cost

Per frame, from zero. Shorter is better

Operations (multiply-adds)

LeanScene
7.2 M
TransNet V2
832 M

Parameters

LeanScene
0.39 M
TransNet V2
7.62 M

How it works

  1. Every frame is decoded and shrunk to 100×100 pixels.
  2. A small network turns each frame into a short description, once.
  3. A second network reads those descriptions in order and labels each frame same, cut or fade.

Both run on your graphics card through WebGPU, or on the CPU where that is missing.

Use it yourself

Python

pip install git+https://github.com/lucasjacks0n/LeanScene
leanscene video.mp4

Node.js

npm install leanscene
npx leanscene video.mp4

Code, weights and training labels on GitHub · npm. The paper is in preparation.

This page reads MP4 and MOV files in a current Chrome, Edge or Safari.