A CPU based video summarization and classification tool. Extracts every movement out of a given video, aggregates the extracted changes into movement layers, which can be exported or further analysed and classified..

Go to file

Askill 8627608d1e added tensorflow to dep tensorflow is optional, and only for the example implementation of the classifier interface		2020-12-26 15:08:34 +01:00
.vscode	contours are getting extracted	2020-09-22 20:25:06 +02:00
Application	added requirements.txt	2020-12-26 14:58:58 +01:00
docs	docs	2020-12-22 13:58:47 +01:00
generate test footage	banished ghosts	2020-12-13 21:36:04 +01:00
.gitignore	.	2020-12-19 14:52:06 +01:00
README.md	more docs	2020-12-22 14:17:21 +01:00
licens.txt	added licens	2020-12-19 14:51:32 +01:00
main.py	removed import of remove class	2020-12-26 15:06:00 +01:00
requirements.txt	added tensorflow to dep	2020-12-26 15:08:34 +01:00

README.md

Video Synopsis and Classification

Example:

What you see above is a 15 second exerpt of a 2 minute overlayed synopsis of a 2.5h video from an on campus web cam.
The synopsis took 40 minutes from start to finish on a 8 core machine and used a maximum of 6Gb of RAM.

However since the contour extraction could be performed on a video stream, the benchmark results show that a single core would be enough to process a video faster than real time.

Benchmark

Below you can find the benchmark results for a 10 minutes clip, with the stacked time per component on the x-axis.
The tests were done on a machine with a Ryzen 3700X with 8 cores 16 threads and 32 Gb of RAM.
On my configuration 1 minutes of of the original Video can be processed in about 20 seconds, the expected processing time is about 1/3 of the orignial video length.

CE = Contour Extractor
LE = LayerFactory
LM = LayerManager
EX = Exporter

notes:

install tensorflow==1.15.0 and tensorflow-gpu==1.15.0, cuda 10.2 and 10.0, copy missing files from 10.0 to 10.2, restart computer, set maximum vram